University of Pennsylvania ScholarlyCommons
IRCS Technical Reports Series Institute for Research in Cognitive Science
August 1998
A Lexicalized Tree Adjoining Grammar for English
The XTAG Research Group University of Pennsylvania
Follow this and additional works at: https://repository.upenn.edu/ircs_reports
Research Group, The XTAG, "A Lexicalized Tree Adjoining Grammar for English" (1998). IRCS Technical Reports Series. 63. https://repository.upenn.edu/ircs_reports/63
University of Pennsylvania Institute for Research in Cognitive Science Technical Report No. IRCS-98-18.
This paper is posted at ScholarlyCommons. https://repository.upenn.edu/ircs_reports/63 For more information, please contact [email protected]. A Lexicalized Tree Adjoining Grammar for English
Abstract This document describes a sizable grammar of English written in the TAG formalism and implemented for use with the XTAG system. This report and the grammar described herein supersedes the TAG grammar described in [XTAG-Group,1995]. The English grammar described in this report is based on the TAG formalism developed in [Joshi et al., 1975], which has been extended to include lexicalization ([Schabes et al., 1998]) and unification-based eaturf e structures ([Vijay-Shanker and Joshi, 1991]). The range of syntactic phenomena that can be handled is large and includes auxiliaries (including inversion), copula, raising and small clause constructions, topicalization, relative clauses, infinitives, gerunds, passives, adjuncts, it-clefts, wh-clefts, PRO constructions, noun-noun modifications, extraposition, determiner sequences, genitives, negation, noun-verb contractions, sentential adjuncts and imperatives. This technical report corresponds to the XTAG Release 8/31/98. The XTAG grammar is continuously updated with the addition of new analyses and modification of old ones, and an online ersionv of this report can be found at the XTAG web page: http://www.cis.upenn.edu/~xtag/.
Comments University of Pennsylvania Institute for Research in Cognitive Science Technical Report No. IRCS-98-18.
This technical report is available at ScholarlyCommons: https://repository.upenn.edu/ircs_reports/63
A Lexicalized Tree Adjoining Grammar for English
The XTAG Research Group
Institute for Research in Cognitive Science
University of Pennsylvania
Walnut St Suite A
Philadelphia PA
http www cis upenn edu xta g
August
Contents
I General Information
Getting Around
Feature Based Lexicalized Tree Adjoining Grammars
TAG formalism
Lexicalization
Uni cation based features
Overview of the XTAG System
System Description
Tree Selection
Tree Database
Tree Grafting
The Grammar DevelopmentEnvironment
Computer Platform
Underview
Sub categorization Frames
Complements and Adjuncts
Non S constituents
Case Assignment
Approaches to Case
Case in GB theory
Minimalism and Case
Case in XTAG
Case Assigners
Prep ositions
Verbs
PRO in a uni cation based framework
II Verb Classes
Where to Find What i
Verb Classes
Intransitive Tnx V
Transitive Tnx Vnx
Ditransitive Tnx Vnx nx
Ditransitive with PP Tnx Vnx pnx
Ditransitive with PP shift Tnx Vnx tonx
Sentential Complement with NP Tnx Vnx s
IntransitiveV erb Particle Tnx Vpl
TransitiveVerb Particle Tnx Vplnx
DitransitiveVerb Particle Tnx Vplnx nx
Intransitive with PP Tnx Vpnx
Predicative Multi word with Verb Prep anchors Tnx VPnx
Sentential Complement Tnx Vs
Intransitive with Adjective Tnx Vax
TransitiveSentential Sub ject Ts Vnx
LightVerbs Tnx lVN
DitransitiveLightVerbs with PP Shift Tnx lVN Pnx
NP It Cleft TItVnx s
PP It Cleft TItVpnx s
Adverb It Cleft TItVad s
ree Tnx Ax Adjective Small Clause T
Adjective Small Clause with Sentential Complement Tnx A s
Adjective Small Clause with Sentential Sub ject Ts Ax
Equative BE Tnx BEnx
NP Small Clause Tnx N
NP Small Clause with Sentential Complement Tnx N s
NP Small Clause with Sentential Sub ject Ts N
PP Small Clause Tnx Pnx
Exhaustive PP Small Clause Tnx Px
PP Small Clause with Sentential Sub ject Ts Pnx
IntransitiveSentential Sub ject Ts V
Sentential Sub ject with to complement Ts Vtonx
PP Small Clause with Adv and Prep anchors Tnx ARBPnx
PP Small Clause with Adj and Prep anchors Tnx APnx
PP Small Clause with Noun and Prep anchors Tnx NPnx
PP Small Clause with Prep anchors Tnx PPnx
PP Small Clause with Prep and Noun anchors Tnx PNaPnx
PP Small Clause with Sen tential Sub ject and Adv and Prep anchors Ts ARBPnx
PP Small Clause with Sentential Sub ject and Adj and Prep anchors Ts APnx
PP Small Clause with Sentential Sub ject and Noun and Prep anchors Ts NPnx
PP Small Clause with Sentential Sub ject and Prep anchors Ts PPnx
PP Small Clause with Sentential Sub ject and Prep and Noun anchors Ts PNaPnx
Predicative Adjective with Sentential Sub ject and Complement Ts A s
Lo cative Small Clause with Ad anchor Tnx nx ARB
Exceptional Case Marking TXnx Vs ii
Idiom with V D and N anchors Tnx VDN
Idiom with V D A and N anchors Tnx VDAN
Idiom with V and N anchors Tnx VN
Idiom with V A and N anchors Tnx VAN
Idiom with V D A N and Prep anchors Tnx VDAN Pnx
Idiom with V A N and Prep anchors Tnx VAN Pnx
Idiom with V N and Prep anchors Tnx VN Pnx
Idiom with V D N and Prep anchors Tnx VDN Pnx
Ergatives
Sentential Sub jects and Sentential Complements
S or VP complements
Complementizers and Emb edded Clauses in English The Data
Features Required
Distribution of Complementizers
Case assignment for and the two to s
Sentential Complements of Verbs
Exceptional Case Marking Verbs
Sentential Sub jects
Nouns and Prep ositions taking Sentential Complements
PRO control
Typ es of control
A feature based analysis of PROcontrol
The nature of the control feature
Long distance transmission of control features
Lo cality constraints on control
Rep orted sp eech
The English Copula Raising Verbs and Small Clauses
Usages of the copula raising verbs and small clauses
Copula
Raising Verbs
Small Clauses
Raising Adjectives
Various Analyses
Main Verb Raising to INFL Small Clause
Auxiliary Null Copula
Auxiliary Predicative Phrase
Auxiliary Small Clause
XTAG analysis
Non predicative BE
Ditransitive constructions and dative shift
It clefts iii
III Sentence Typ es
Passives
Extraction
Topicalization and the value of the inv feature
Extracted sub jects
Wh moved NP complement
Wh moved ob ject of a P
Wh moved PP
Wh moved S complement
Wh moved Adjective complement
Relative Clauses
Complementizers and clauses
Further constraints on the null Comp
C
Reduced Relatives
Restrictivevs Non restrictive relatives
External syntax
Other Issues
Interaction with adjoined Comps
Adjunction on PRO
Adjunct relative clauses
ECM
Cases not handled
Partial treatment of free relatives
Adjunct P stranding
Overgeneration
how as wh NP
for trace e ects
Internal head constraint
Overt Comp constraintonstacked relatives
Adjunct Clauses
Multi word Sub ordinating Conjunctions
Bare Adjunct Clauses
Discourse Conjunction
Imp eratives
Gerund NP s
Determiner Gerunds
NP Gerunds
Gerund Passives iv
IV Other Constructions
Determiners and Noun Phrases
The Wh Feature
Multi word Determiners
Genitive Constructions
Partitive Constructions
Adverbs Noun Phrases and Determiners
Mo di ers
Adjectives
Noun Noun Mo di ers
Time Noun Phrases
Prep ositions
Adverbs
Lo cativeAdverbial Phrases
Auxiliaries
Non inverted sentences
Inverted Sentences
Do Supp ort
In negated sentences
In inverted yes no questions
In nitives
Semi Auxiliaries
Marginal Mo dal dare
Other semi auxiliaries
Other Issues
Conjunction
Intro duction
Adjective Adverb Prep osition and PP Conjunction
Noun Phrase and Noun Conjunction
Determiner Conjunction
Sentential Conjunction
Comma as a conjunction
But not not but and not and not
To as a Conjunction
Predicative Co ordination
Pseudo co ordination
Comparatives
Intro duction
Metalinguistic Comparatives
Prop ositional Comparatives
Nominal Comparatives v
Adjectival Comparatives
Adverbial Comparatives
Future Work
Punctuation Marks
App ositives parentheticals and vo catives
nxPUnxPU
nPUnxPU
nxPUnx
PUpxPUvx
puARBpuvx
sPUnx
nxPUs
Bracketing punctuation
Simple bracketing
sPUsPU
Punctuation trees containing no lexical material
PU
PUs
sPUs
sPU
vPU
pPU
Other trees
spuARB
spuPnx
nxPUa
V App endices
A Future Work
A Adjective ordering
A More work on Determiners
A ing adjectives
A Verb selectional restrictions
A Thematic Roles
B Metarules
B Intro duction
B The de nition of a metarule in XTAG
ariable instantiation and matches B No de names v
B Structural Matching
B Output Generation
B Feature Matching
B Examples vi
B The Access to the Metarules through the XTAGInterface
C Lexical Organization
C Intro duction
C System Overview
C Sub categorization frames
C Blo cks
C Lexical Redistribution Rules LRRs
C Tree generation
C Implementation
C Generating grammars
C Summary
D Tree Naming conventions
D Tree Families
D Trees within tree families
D Assorted Initial Trees
D Assorted Auxiliary Trees
D Relative Clause Trees
E Features
E Agreement
E Agreement and Movement
E Case
E ECM
E Agreement and Case
E Extraction and Inversion
E Inversion Part
ersion Part E Inv
E Clause Typ e
E Auxiliary Selection
E Relative Clauses
E Complementizer Selection
E Verbs with ob ject sentential complements
E Verbs with sentential sub jects
E That trace and for trace e ects
E Determiner ordering
E Punctuation
E Conjunction
E Comparatives
E Control
E Other Features vii
F Evaluation and Results
F Parsing Corp ora
F TSNLP
F Chunking and Dep endencies in XTAG Derivations
F Comparison with IBM
F Comparison with Alvey
F Comparison with CLARE viii
List of Figures
Elementary trees in TAG
Substitution in TAG
Adjunction in TAG
Lexicalized Elementary trees
Substitution in FB LTAG
Adjunction in FB LTAG
Lexicalized Elementary Trees with Features
XTAG system diagram
Output Structures from the Parser
Interfaces database
XTAGInterface
Di eren t sub categorization frames for the verb buy
Trees illustrating the di erence b etween Complements and Adjuncts
Lexicalized NP trees with case markings
Assigning case in prep ositional phrases
Case assignment to NP arguments
Assigning case according to verb form
Prop er case assignment with auxiliary verbs
Declarative IntransitiveTree nx V
Declarative TransitiveTree nx Vnx
Declarative DitransitiveTree nx Vnx nx
Declarative Ditransitive with PP Tree nx Vnx pnx
nx Vnx nx b Declarative Ditransitive with PP shift Trees nx Vnx Pnx a and
Declarative Sentential Complement with NP Tree nx Vnx s
Declarative IntransitiveVerb Particle Tree nx Vpl
Declarative TransitiveVerb Particle Tree nx Vplnx a and nx Vnx pl b
Declarative DitransitiveVerb Particle Tree nx Vplnx nx a nx Vnx plnx b
and nx Vnx nx pl c
Declarative IntransitivewithPPTree nx Vpnx
Declarative PP ComplementTree nx VPnx
Declarative Sentential ComplementTree nx Vs
Declarative Intransitive with AdjectiveT ree nx Vax
Declarative Sentential Sub ject Tree s Vnx ix
Declarative LightVerb Tree nx lVN
Declarative LightVerbs with PP Tree nx lVN Pnx a nx lVnx N b
Declarative NP It Cleft Tree ItVpnx s
Declarative PP It Cleft Tree ItVnx s
Declarative Adverb It Cleft Tree ItVad s
Declarative Adjective Small Clause Tree nx Ax
Declarative Adjective Small Clause with Sentential ComplementTree nx A s
Declarative Adjective Small Clause with Sentential Sub ject Tree s Ax
Declarative Equative BE Tree nx BEnx
Declarative NP Small Clause Trees nx N
Declarative NP with Sentential Complement Small Clause Tree nx N s
Declarative NP Small Clause with Sentential Sub ject Tree s N
Declarative PP Small Clause Tree nx Pnx
Declarative Exhaustive PP Small Clause Tree nx Px
Declarative PP Small Clause with Sentential Sub ject Tree s Pnx
Declarative IntransitiveSentential Sub ject Tree s V
Sentential Sub ject Tree with to complement s Vtonx
Declarative PP Small Clause tree with two word prep osition where the rst word
is an adverb and the second word is a prep osition nx ARBPnx
Declarative PP Small Clause tree with two word prep osition where the rst word
is an adjective and the second word is a prep osition nx APnx
Declarative PP Small Clause tree with two word prep osition where the rst word
is a noun and the second word is a prep osition nx NPnx
Declarative PP Small Clause tree with two word prep osition where b oth words
are prep ositions nx PPnx
Declarative PP Small Clause tree with three word prep osition where the middle
noun is marked for null adjunction nx PNaPnx
Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where the rst word is an adverb and the second word is a prep osition
s ARBPnx
o word prep o Declarative PP Small Clause with Sentential Sub ject Tree with tw
sition where the rst word is an adjective and the second word is a prep osition
s APnx
Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where the rst word is a noun and the second word is a prep osition
s NPnx
Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where b oth words are prep ositions s PPnx
Declarative PP Small Clause with Sentential Sub ject Tree with three word
prep osition where the middle noun is marked for null adjunction s PNaPnx
Predicative Adjective with Sentential Sub ject and Complement s A s
Declarative Lo cativeAdverbial Small Clause Tree nx nx ARB
Wh moved Lo cative Small Clause Tree W nx nx ARB
ECM Tree Xnx Vs
Declarative Transitive Idiom Tree nx VDN x
Declarative Idiom with V D A and N Anchors Tree nx VDAN
Declarative Idiom with V and N Anchors Tree nx VN
Declarative Idiom with V A and N Anchors Tree nx VAN
Declarative Idiom with V D A N and Prep Anchors Tree nx VDAN Pnx
Declarative Idiom with V A N and Prep Anchors Tree nx VAN Pnx
Declarative Idiom with V N and Prep Anchors Tree nx VN Pnx
e Idiom with V D N and Prep Anchors Tree nx VDN Pnx Declarativ
ErgativeTree Enx V
Tree COMPs anchored by that
Sentential complement tree nx Vs
Trees for The emu thinks that the aardvark smel ls terrible
Tree for Who smel ls terrible
ECM tree Xnx Vs
Sample ECM parse
ECM passive
Comparison of assign comp values for sentential sub jects s Vnx a and
sentential complements nx Vs b
Sample trees for prep osition Pss a and noun NXNs b taking sentential
complements
Tree for persuaded
Tree for leave
Derived tree for Srini persuaded Mickey to leave
Derivation tree for Srini persuaded Mickey to leave
Predicative trees nx N a nx Ax b and nx Pnx c
Copula auxiliary tree Vvx
PredicativeAPtreewithfeatures nx Ax
Consider tree for emb edded small clauses
Raising verb with exp eriencer tree Vpxvx
Raising adjective tree Vvx adj
Equativ e BE trees nx BEnx a and Invnx BEnx b
Dative shift trees nx Vnx Pnx a and nx Vnx nx b
It cleft with PP clefted element ItVpnx s a and InvItVpnx s b
Passive trees in the Sentential Complement with NP tree family nx Vs a
nx Vbynx s b and nx Vs bynx c
e tree with ob ject extraction W nx Vnx Transitiv
Intransitive tree with sub ject extraction W nx V
Ditransitive trees with direct ob ject W nx Vnx nx a and indirect ob ject
extraction W nx Vnx nx b
Ditransitive with PP tree with the ob ject of the PP extracted W nx Vnx pnx
Ditransitive with PP with PP extraction tree pW nx Vnx pnx xi
Predicative Adjective tree with extracted adjective WA nx Vax
Relative clause trees in the transitive tree family Nc nx Vnx a and N nx Vnx
b
Adjunct relative clause tree with PP pied piping in the transitive tree family
Npxnx Vnx
Determiner tree with rel clause feature Dnx
Auxiliary Trees for Sub ordinating Conjunctions
Trees Anchored by Sub ordinating Conjunctions vxPARBPs and vxParbPs
Sample Participial Adjuncts
Example of discourse conjunction from Seuss The Lorax
Transitive imp erative tree Inx Vnx
Determiner Gerund tree from the transitive tree family Dnx Vnx
NP Gerund tree from the transitive tree family Gnx Vnx
Passive Gerund trees from the transitive tree family Gnx Vbynx a and
Gnx V b
NP Tree
Determiner Trees with Features
Multi word Determiner tree DDnx
Genitive Determiner Tree
Genitive NP tree for substitution DnxG
Partitive Determiner Tree
a Adverb mo difying a determiner b Adverb mo difying a noun phrase
Standard Tree for Adjective mo difying a Noun An
Multiple adjectives mo difying a noun
Noun noun comp ounding tree Nn not all features displayed
Time Phrase Mo di er trees Ns Nvx vxN nxN
Time NPs with and without a determiner
Time NP trees Two di erent attachments
Time NPs in di erent p ositions vxN nxN and Ns
Time NPs Derived tree and Derivation Nvx p osition
Selected Prep ositional Phrase Mo di er trees Pss nxPnx vxP and vxPPnx
Adverb Trees for pre mo di cation of S ARBs a and p ost mo di cation of a
VP vxARB b
Derived tree for How did you fal l
Complex adverb phrase mo di er ARBarbs
Selected Focus and Multi word Adverb Mo di er trees ARBnx PARBd and
PaPd
Selected Multi word Adverb Mo di er trees for adverbs like sort of kind of
NPax NPvx vxNP
Selected Multi word Adverb Mo di er trees for adverbs like alittle abit vxDA
DAax DNpx xii
Lo cative Mo di er Trees nxnxARB nxARB
Lo cative Phrases featuring NP and Adverb Degree Sp eci cations
Auxiliary verb tree for non inverted sentences Vvx
Auxiliary trees for The music should have been being played
The music should have been being played
Trees for auxiliary verb inversion Vs a and Vvx b
wil l John buy a backpack
Tree for adjective conjunction a CONJa and a resulting parse tree
Tree for NP conjunction CONJnx CONJnx and a resulting parse tree
Tree for determiner conjunction d CONJd ps
Tree for sentential conjunction s CONJs
a CONJa a anchored by comma and b anchored by and
Tree for conjunction with but not px CONJARBpx
Tree for conjunction with not but ARBnx CONJnx
Example of conjunction with to
Co ordination schema
An example of the conjoin op eration f g denotes a shared dep endency
Co ordination as adjunction
Tree for Metalinguistic Adjective Comparative ARBaPa
Tree for Adjective Extreme Comparative ARBPa
Nominal comparative trees
Tree for Lone Comparatives CARB
Comparative conjunctions
Comparative conjunctions
Adjunction of nxPnx to NP mo di ed by comparative adjective
Elliptical adjectival comparative trees
Comparativized adjective triggering CnxPnx
Adjunction of axPnx to comparative adjective
Adverbial comparative trees
The nxPUnxPU tree anchored byparentheses
An N level mo di er using the nPUnx tree
The derived trees for an NP with a a p eripheral dash separated app ositive
and b an NP colon expansion uttered by the Mouse in Alice s Adventures in
Wonderland
The PUpxPUvx tree anchored by commas
Tree illustrating the use of PUpxPUvx
A tree illustrating the use of sPUnx for a colon expansion attached at S
PUsPU anchored by parentheses and in a derivation along with PUnxPU
PUs with features displayed
sPUs with features displayed xiii
B Metarule for wh movement of sub ject
B Metarule for wh movement of ob ject
B Metarule for general wh movementofanNP
B Application of wh movement rule to Tnx Vnx Pnx
B Parallel application of metarules
B Sequential application of metarules
B Cumulative application of metarules
C Lexical Organization System Overview
C Some sub categorization blo cks
C Transformation blo cks for extraction
C Elementary trees generated from combining blo cks
C Partial inheritance lattice in English
C Implementation of the system
C Interface for creating a grammar
C Part of the Interface for creating blo cks xiv
Abstract
This do cument describ es a sizable grammar of English written in the TAG formalism and
implemented for use with the XTAG system This rep ort and the grammar describ ed herein
sup ersedes the TAG grammar describ ed in XTAG Group The English grammar de
scrib ed in this rep ort is based on the TAG formalism develop ed in Joshi et al which has
been extended to include lexicalization Schab es et al and uni cation based feature
structures Vijay Shanker and Joshi The range of syntactic phenomena that can be
handled is large and includes auxiliaries including inversion copula raising and small clause
constructions topicalization relative clauses in nitives gerunds passives adjuncts it clefts
noun noun mo di cations extrap osition determiner sequences wh clefts PRO constructions
genitives negation noun verb contractions sentential adjuncts and imp eratives This techni
cal rep ort corresp onds to the XTAG Release The XTAG grammar is continuously
up dated with the addition of new analyses and mo di cation of old ones and an online version
of this rep ort can b e found at the XTAGweb page http www cis upenn edu xtag
Acknowledgements
We are immensely grateful to Aravind Joshi for supp orting this pro ject
The following p eople have contributed to the development of grammars in the pro ject Anne
Ab eille Jason Baldridge Ra jesh Bhatt Kathleen Bishop Raman Chandrasekar Sharon Cote
Beatrice Daille Christine Doran Dania Egedi Tim Farrington Jason Frank Caroline Heyco ck
Beth Ann Ho ckey Roumyana Izvorski Karin Kipp er Daniel Karp Seth Kulick Young Suk
Lee Heather Matayek Patrick Martin Megan Moser Sabine Petillon Rashmi Prasad Laura
Siegel Yves Schab es Victoria Tredinnick and Ra aella Zanuttini
The XTAG system has been develop ed by Tilman Becker Richard Billington Andrew
Chalnick Dania Egedi Devtosh Khare Alb ert Lee David Magerman Alex Mallet Patrick
Paroub ek Rich Pito Gilles Prigent Carlos Prolo Ano op Sarkar Yves Schab es William
Schuler B Srinivas Fei Xia Yuji Yoshiie and Martin Zaidel
We would also like to thank Michael Hegarty Lauri Karttunen Anthony Kro ch Mitchell
Marcus Martha Palmer Owen Rambow Philip Resnik Beatrice Santorini and Mark Steedman
In addition Je Aaronson Douglas DeCarlo Mark Jason Dominus Mark Foster Gaylord
Holder David Magerman Ken Noble Steven Shapiro and Ira Winston have provided technical
supp ort Adminstrative supp ort was provided by Susan Deysher Carolyn Elken Jo di Kerp er
Christine Sandy and Trisha Yannuzzi
This work was partially supp orted by NSF Grant SBR and ARO GrantDAAH G
Part I
General Information
Chapter
Getting Around
This technical rep ort presents the English XTAG grammar as implemented by the XTAG
Research Group at the UniversityofPennsylvania The technical rep ort is organized into four
parts plus a set of app endices Part contains general information ab out the XTAG system
and some of the underlying mechanisms that help shap e the grammar Chapter contains
an intro duction to the formalism b ehind the grammar and parser while Chapter contains
information ab out the entire XTAG system Linguists interested solely in the grammar of the
XTAG system may safely skip Chapters and Chapter contains information on some of
the linguistic principles that underlie the XTAG grammar including the distinction between
complements and adjuncts and how case is handled
The actual description of the grammar b egins with Part and is contained in the following
three parts Parts and contains information on the verb classes and the typ es of trees
allowed within the verb classes resp ectively while Part contains information on trees not
included in the verb classes e g NP s PP s various mo di ers etc Chapter of Part
overview of the verb classes and tree typ es by contains a table that attempts to provide an
providing a graphical indication of whichtreetyp es are allowed in whichverb classes This has
been cross indexed to tree gures shown in the tech rep ort Chapter contains an overview
of all of the verb classes in the XTAG grammar The rest of Part contains more details on
several of the more interesting verb classes including ergatives sentential sub jects sentential
complements small classes ditransitives and it clefts
Part contains information on some of the tree t yp es that are available within the verb
classes These tree typ es corresp ond to what would be transformations in a movement based
approach Not all of these typ es of trees are contained in all of the verb classes The table
previously mentioned in Part contains a list of the tree typ es and indicates which verb
classes each o ccurs in
Part fo cuses on the non verb class trees in the grammar NP s and determiners are
presented in Chapter while the various mo di er trees are presented in Chapter Auxiliary
h are classed separate from the verb classes are presented in Chapter while verbs whic
certain typ es of conjunction are shown in Chapter The XTAG treatment of comparatives
is presented in Chapter and our treatment of punctuation is discussed in Chapter
Throughout the technical rep ort mention is o ccasionally made of changes or analyses that
we hop e to incorp orate in the future App endix A details a list of these and other future
work The app endices also contain information on some of the nitty gritty details of the
CHAPTER GETTING AROUND
XTAG grammar including a system of metarules which can b e used for grammar development
and maintenance in App endix B a system for the organization of the grammar in terms of
an inheritance hierarchy is in App endix C the tree naming conventions used in XTAG are
explained in detail in App endix D and a comprehensive list of the features used in the grammar
is given in App endix E App endix F contains an evaluation of the XTAG grammar including
comparisons with other wide coverage grammars
Chapter
Feature Based Lexicalized Tree
Adjoining Grammars
The English grammar describ ed in this rep ort is based on the TAG formalism Joshi et
al which has been extended to include lexicalization Schab es et al and
uni cation based feature structures Vijay Shanker and Joshi Tree Adjoining Lan
guages TALs fall into the class of mildly context sensitive languages and as such are more
powerful than context free languages The TAG formalism in general and lexicalized TAGs
in particular are well suited for linguistic applications As rst shown by Joshi and
Kro ch and Joshi the prop erties of TAGs p ermit us to encapsulate diverse syntactic
phenomena in a very natural way For example TAG s extended domain of lo cality and its
factoring of recursion from lo cal dep endencies lead among other things to a lo calization of
so called unb ounded dep endencies
TAG formalism
The primitive elements of the standard TAG formalism are known as elementary trees Ele
mentary trees areoftwotyp es initial trees and auxiliary trees see Figure In describing
natural language initial trees are minimal linguistic structures that contain no recursion
i e trees containing the phrasal structure of simple sentences NP s PP s and so forth Initial
trees are characterized by the following all internal no des are lab eled by non terminals
all leaf no des are lab eled by terminals or by non terminal no des marked for substitution An
initial tree is called an X typ e initial tree if its ro ot is lab eled with typ e X
Recursive structures are represented by auxiliary trees which represent constituents that
are adjuncts to basic structures e g adverbials Auxiliary trees are characterized as follows
all internal no des are lab eled by non terminals all leaf no des are lab eled by terminals
or by non terminal no des marked for substitution except for exactly one non terminal no de
which can only be used to adjoin the tree to another no de the fo ot called the fo ot no de
no de has the same lab el as the ro ot no de of the tree
Anull adjunction constraint NA is systematically put on the fo ot no de of an auxiliary tree This disallows
adjunction of a tree onto the fo ot no de itself
CHAPTER FEATURE BASED LEXICALIZED TREE ADJOINING GRAMMARS
Initial Tree: Auxiliary Tree:
X X
X
Figure Elementary trees in TAG
There are two op erations de ned in the TAG formalism substitution and adjunction In
the substitution op eration the ro ot no de on an initial tree is merged into a non terminal leaf
no de marked for substitution in another initial tree pro ducing a new tree The ro ot no de and
the substitution no de must have the same name Figure shows two initial trees and the tree
resulting from the substitution of one tree into the other
Y X X 2 => Y Y
1 2
Figure Substitution in TAG
In an adjunction op eration an auxiliary tree is grafted onto a non terminal no de anywhere
in an initial tree The ro ot and fo ot no des of the auxiliary tree must match the no de at which
the auxiliary tree adjoins Figure shows an auxiliary tree and an initial tree and the tree
resulting from an adjunction op eration
A TAG G is a collection of nite initial trees I and auxiliary trees A The tree set of
a TAG G T G is de ned to be the set of all derived trees starting from S typ e initial trees
in I whose frontier consists of terminal no des all substitution no des having b een lled The
string language generated by a TAG L G is de ned to b e the set of all terminal strings
G on the frontier of the trees in T
Technically substitution is a sp ecialized version of adjunction but it is useful to make a distinction b etween
the two X X Y 2 => Y Y 1 2 Y 3
Y
3
Figure Adjunction in TAG
Lexicalization
Lexicalized grammars systematically asso ciate each elementary structure with a lexical anchor
This means that in each structure there is a lexical item that is realized It do es not mean simply
adding feature structures such as head and uni cation equations to the rules of the formalism
These resultant elementary structures sp ecify extended domains of lo cality as compared to
CFGs over which constraints can b e stated
Following Schab es et al we say that a grammar is lexicalized if it consists of
a nite set of structures each asso ciated with a lexical item and an op eration or op erations
for comp osing the structures Each lexical item will b e called the anchor of the corresp onding
structure which de nes the domain of lo calityover which constraints are sp eci ed Note then
that constraints are lo cal with resp ect to their anchor
Not every grammar is in a lexicalized form In the pro cess of lexicalizing a grammar the
lexicalized grammar is required to b e strongly equivalent to the original grammar i e it must
pro duce not only the same language but the same structures or tree set as well
VPr
Sr VP* PP NP NA NP NP0↓ VP
P NP↓ N V N
John walked to Philadelphia
a b c d
Figure Lexicalized Elementary trees
Notice the similarity of the de nition of a lexicalized grammar with the o line parsability constraint Kaplan
and Bresnan As consequences of our de nition each structure has at least one lexical item its anchor
attached to it and all sentences are nitely ambiguous
CHAPTER FEATURE BASED LEXICALIZED TREE ADJOINING GRAMMARS
In Figure which shows sample initial and auxiliary trees substitution sites are marked
by a and fo ot no des are marked by an This notation is standard and is followed in the
rest of this rep ort
Uni cation based features
In a uni cation framework a feature structure is asso ciated with each no de in an elementary
tree This feature structure contains information ab out how the no de interacts with other
no des in the tree It consists of a top part which generally contains information relating to the
sup erno de and a b ottom part which generally contains information relating to the subno de
Substitution no des however have only the top features since the tree substituting in logically carries the b ottom features
tr Y X X br => t t U tr Y Y
br
Figure Substitution in FB LTAG
The notions of substitution and adjunction must b e augmented to t within this new frame
work The feature structure of a new no de created by substitution inherits the union of the
features of the original no des The top feature of the new no de is the union of the top features
of the two original no des while the b ottom feature of the new no de is simply the b ottom feature
of the top no de of the substituting tree since the substitution no de has no bottom feature
Figure shows this more clearly
Adjunction is only slightly more complicated The no de b eing adjoined into splits and its
top feature uni es with the top feature of the ro ot adjoining no de while its b ottom feature uni
es with the b ottom feature of the fo ot adjoining no de Again this is easier shown graphically
as in Figure
The emb edding of the TAG formalism in a uni cation framework allows us to dynamically
sp ecify lo cal constraints that would have otherwise had to have been made statically within
the trees Constraints that verbs make on their complements for instance can b e implemented
notions of Obligatory and Selective Adjunction crucial through the feature structures The
abbreviations in the gure t top feature structure tr top feature structure of the ro ot br b ottom feature
structure of the ro ot U uni cation
abbreviations in the gure t top feature structure b b ottom feature structure tr top feature structure
of the ro ot br b ottom feature structure of the ro ot tf top feature structure of the fo ot bf b ottom feature
structure of the fo ot U uni cation tr X Y br X
t t U tr Y b Y > tf br Y* bf
tf Y
b U bf
Figure Adjunction in FB LTAG
to the formation of lexicalized grammars can also be handled through the use of features
Perhaps more imp ortant to developing a grammar though is that the trees can serve as a
schemata to b e instantiated with lexical sp eci c features when an anchor is asso ciated with the
tree To illustrate this Figure shows the same tree lexicalized with two di erentverbs each
of which instantiates the features of the tree according to its lexical selectional restrictions
In Figure the lexical item thinks takes an indicative sentential complement as in the
sentence John thinks that Mary loves Sal ly Want takes a sentential complementaswell but an
in nitiveone asinJohn wants to love Mary This distinction is easily captured in the features
and passed to other no des to constrain which trees this tree can adjoin into b oth cutting down
the numb er of separate trees needed and enforcing conceptual Selective Adjunctions SA
The remaining constraint Null Adjunction NA must still b e sp eci ed directly on a no de
CHAPTER FEATURE BASED LEXICALIZED TREE ADJOINING GRAMMARS
Sr assign-comp : inf_nil/ind_nil Sr assign-comp : inf_nil/ind_nil displ-const : set1 : - displ-const : set1 : - tense : <1> tense : <1> assign-case : <2> assign-case : <2> agr : <3> agr : <3> assign-comp : <4> assign-comp : <4> mode : <5> mode : <5> inv : - inv : - comp : nil comp : nil displ-const : set1 : <6> displ-const : set1 : <6> wh : <7> - wh : <7> - extracted : - extracted : -
NP0↓ case : <2> VP assign-case : <2> NP0↓ case : <2> VP assign-case : <2> agr : <3> agr : <3> agr : <3> agr : <3> wh : <7> tense : <1> wh : <7> tense : <1> assign-comp : <4> assign-comp : <4> mode : <5> mode : <5> displ-const : set1 : <6> displ-const : set1 : <6> mainv : <8> mainv : <8> tense : <9> tense : <9> mode : <10> mode : <10> assign-comp : <11> assign-comp : <11> assign-case : <12> assign-case : <12> agr : <13> agr : <13> passive : <14> - passive : <14> - displ-const : set1 : - displ-const : set1 : -
V mainv : <8> S1* displ-const : set1 : - V mainv : <8> S1* displ-const : set1 : - tense : <9> assign-comp : inf_nil/ind_nil tense : <9> assign-comp : inf_nil/ind_nil mode : <10> inv : - mode : <10> inv : - assign-comp : <11> comp : that/whether/if/nil assign-comp : <11> comp : whether/for/nil assign-case : <12> mode : ind/sbjnct assign-case : <12> mode : inf agr : <13> agr : <13> passive : <14> passive : <14> mode : ind mode : ind tense : pres tense : pres mainv : - mainv : - assign-comp : ind_nil/that/rel/if/whether assign-comp : ind_nil/that/rel/if/whether assign-case : nom assign-case : nom agr : 3rdsing : + agr : 3rdsing : + num : sing num : sing pers : 3 pers : 3
thinks wants
think tree want tree
Figure Lexicalized Elementary Trees with Features
Chapter
Overview of the XTAG System
This section fo cuses on the various comp onents that comprise the parser and English grammar
in the XTAG system Persons interested only in the linguistic analyses in the grammar may
skip this section without loss of continuity although a quick glance at the tagset used in XTAG
and the set of non terminal lab els used will b e useful We may o ccasionally refer back to the
various comp onents mentioned in this section
System Description
Figure shows the overall ow of the system when parsing a sentence a summary of each
comp onent is presented in Table At the heart of the system is a parser for lexicalized
TAGs Schab es and Joshi Schab es which pro duces all legitimate parses for the
sentence The parser has two phases Tree Selection and Tree Grafting
Tree Selection
Since we are working with lexicalized TAGs each word in the sentence selects at least one
tree The advantage of a lexicalized formalism likeLTAGs is that rather than parsing with all
the trees in the grammar we can parse with only the trees selected bythe words in the input
sentence
several steps Each In the XTAG system the selection of trees by the words is done in
step attempts to reduce ambiguity i e reduce the numb er of trees selected by the words in the
sentence
Morphological Analysis and POS Tagging The input sentence is rst submitted to the
Morphological Analyzer and the Tagger The morphological analyzer Karp et al
consists of a disk based database a compiled version of the derivational rules
which is used to map an in ected word into its stem part of sp eech and feature equations
corresp onding to in ectional information These features are inserted at the anchor no de
of the tree eventually selected by the stem The POS Tagger can be disabled in which
case only information from the morphological analyzer is used The morphology data
and Oxford was originally extracted from the Collins English Dictionary Hanks
Advanced Learner s Dictionary Hornby available through ACL DCI Lib erman
and then cleaned up and augmented by hand Karp et al
CHAPTER OVERVIEW OF THE XTAG SYSTEM
Input Sentence
Morph Analyzer Tagger
Morph DB Lex Prob DB P.O.S Blender
Stat DB Tree Selection Syn DB Trees DB Parser
Tree Grafting
Derivation Structure
Figure Overview of XTAG system
POS Blender The output from the morphological analyzer and the POS tagger go into the
POS Blender whichusesthe output of the POS tagger as a lter on the output of the
morphological analyzer Anywords that are not found in the morphological database are
assigned the POS given by the tagger
Syntactic Database The syntactic database contains the mapping b etween particular stem s
and the tree templates or tree families stored in the Tree Database see Table The
syntactic database also contains a list of feature equations that capture lexical idiosyn
crasies The output of the POS Blender is used to search the Syntactic Database to
pro duce a set of lexicalized trees with the feature equations asso ciated with the word s
in the syntactic database uni ed with the feature equations asso ciated with the trees
Note that the features in the syntactic database can b e assigned to any no de in the tree
and not just to the anchor no de The syntactic database entries were originally extracted
from the Oxford Advanced Learner s Dictionary Hornby and Oxford Dictionary
for Contemp orary Idiomatic English Cowie and Mackin available through ACL
DCI Lib erman and then mo di ed and augmented by hand Egedi and Martin
There are more than syntactic database entries Selected entries from
this database are shown in Table
Default Assignment For words that are not found in the syntactic database default trees
and tree families are assigned based on their POS tag
Filters Some of the lexicalized trees chosen in previous stages can be eliminated in order to
h eliminate reduce ambiguity Two metho ds are currently used structural lters whic
trees whichhave imp ossible spans over the input sentence and a statistical lter based on
This numb er do es not include trees assigned by default based on the part of sp eech of the word
Comp onent Details
Morphological Consists of approximately in ected items
Analyzer and derived from over stems
Morph Database Entries are indexed on the in ected form and return
the ro ot form POS and in ectional information
POS Tagger Wall Street Journal trained trigram tagger Church
and Lex Prob extended to output N b est POS sequences
Database So ong and Huang Decreases the time to parse
asentence byanaverage of
Syntactic More than entries
Database Eachentry consists of the unin ected form of the word
its POS the list of trees or tree families asso ciated with
the word and a list of feature equations that capture
lexical idiosyncrasies
Tree Database trees divided into tree families and individual
trees Tree families represent sub categorization frames
the trees in a tree family would b e related to eachother
transformationally in a movement based approach
X Interface Menu based facility for creating and mo difying tree les
User controlled parser parameters parser s start category
enable disable retry on failure for POS tagger
Storage retrieval facilities for elementary and parsed trees
Graphical displays of tree and feature data structures
Hand combination of trees by adjunction or substitution
for grammar development
Abilitytomanually assign POS tag
and or Sup ertag b efore parsing
Table System Summary
unigram probabilities of non lexicalized trees from a hand corrected set of approximately
parsed sentences These metho ds sp eed the runtime by approximately
Sup ertagging Before parsing one can avail of an optional step of supertagging the sentence
This step uses statistical disambiguation to assign a unique elementary tree or supertag
to each word in the sentence These assignments can then be hand corrected These
sup ertags are used as a lter on the tree assignments made so far More information on
sup ertagging can b e found in Srinivas a Srinivas b
Tree Database
The Tree Database contains the tree templates that are lexicalized by following the various
steps given ab ove The lexical items are inserted into distinguished no des in the tree template
to the called the anchor nodes The part of sp eech of each word in the sentence corresp onds
lab el of the anchor no de of the trees Hence the tagset used by the POS Tagger corresp onds
exactly to the lab els of the anchor no des in the trees The tagset used in the XTAG system is
given in Table The tree templates are sub divided into tree families for verbs and other
CHAPTER OVERVIEW OF THE XTAG SYSTEM
INDEX porousness ENTRY porousness POS N
TREES BNXN BN CNn
FEATURES N card N const N decreas N definite N gen
N quan N refl
INDEX coo ENTRY coo POS V FAMILY Tnx V
INDEX engross ENTRY engross POS V FAMILY Tnx Vnx
FEATURES TRANS
INDEX forbear ENTRY forbear POS V FAMILY Tnx Vs
FEATURES S WH S inf for nil
INDEX have ENTRY have POS V ENTRY out POS PL
FAMILY Tnx Vplnx
Table Example Syntactic Database Entries
predicates and tree les which are simply collections of trees for lexical items like prep ositions
determiners etc
Tree Grafting
Once a particular set of lexicalized trees for the sentence have been selected XTAG uses an
Earley style predictive left to right parsing algorithm for LTAGs Schab es and Joshi
Schab es to nd all derivations for the sentence The derivation trees and the asso ciated
derived trees can b e viewed using the X interface see Table The X interface can also be
used to save particular derivations to disk
The output of the parser for the sentence Ihadamap yesterday is illustrated in Figure
The parse tree represents the surface constituent structure while the derivation tree represents
the derivation history of the parse The no des of the derivation tree are the tree names anchored
by the lexical items The comp osition op eration is indicated by the nature of the arcs a dashed
line is used for substitution and a b old line for adjunction The numb er b eside each tree name
is the address of the no de at whichthe op eration to ok place The derivation tree can also be
interpreted as a dep endency graph with unlab eled arcs b etween words of the sentence
The Grammar Development Environment
Working with and developing a large grammar is a challenging pro cess and the imp ortance of
having good visualization to ols cannot be over emphasized Currently the XTAG system has
The nonterminals in the tree database are A AP Ad AdvP Comp Conj D N NP P PP Punct S
V VP
The feature structures asso ciated with each note of the parse tree are not shown here
App endix D explains the conventions used in naming the trees Sr
NP VPr
N VP PP NA αnx0Vnx1[had]
I V NPr P NPr αNXN[I] (1) βvxPnx[on] (2) αNXN[map] (2.2) had D NPf on D NPf NA NA αNXN[desk] (2.2) βDnx[a] (0) a N my N
map desk βDnx[my] (0)
Parse Tree Derivation Tree
Figure Output Structures from the Parser
Part of Sp eech Description
A Adjective
Ad Adverb
Comp Complementizer
D Determiner
G GenitiveNoun
I Interjection
N Noun
P Prep osition
PL Particle
Punct Punctuation
V Verb
Table XTAG tagset
X windows based to ols for viewing and up dating the morphological and syntactic databases
Karp et al Egedi and Martin These are available in b oth ASCI I and binary
enco ded database format The ASCI I format is well suited for various UNIX utilities awk
sed grep while the database format is used for fast access during program execution However
even the ASCI I formatted representation is not well suited for human readability An X
windows interface for the databases allows users to easily examine them Searching for sp eci c
information on certain elds of the syntactic database is also available Also the interface allows
a user to insert delete and up date any information in the databases Figure a shows the
shows the interface for the syntactic interface for the morphology database and Figure b
database
XTAG also has a parsing and grammar developmentinterface Paroub ek et al This
CHAPTER OVERVIEW OF THE XTAG SYSTEM
a Morphology database b Syntactic database
Figure Interfaces to the database maintenance to ols
interface includes a tree editor the abilitytovary parameters in the parser work with multiple
grammars and or parsers and use metarules for more e cient tree editing and construction
Becker The interface is shown in Figure It has the following features
Menu based facility for creating and mo difying tree les and loading grammar les
User controlled parser parameters including the ro ot category main S emb edded S NP
etc and the use of the tagger on o retry on failure
Storage retrieval facilities for elementary and parsed trees
The pro duction of p ostscript les corresp onding to elementary and parsed trees
Graphical displays of tree and feature data structures including a scroll web for large
tree structures
Mouse based tree editor for creating and mo difying trees and feature structures
Hand combination of trees by adjunction or substitution for use in diagnosing grammar
problems
tree based transfor Metarule to ol for automatic aid to the generation of trees by using
mation rules
Figure Interface to the XTAG system
Computer Platform
XTAG was develop ed on the Sun SPARC station series It has been tested on various Sun
platforms including Ultra Ultra Enterprise XTAG is freely available from the XTAG web
page at http www cis upenn edu xtag It requires MB of disk space once all binaries
and databases are created after the install XTAG requires the following software to run
A machine running UNIX and X R or higher Previous releases of Xwillnot work
X R is free software which usually comes bundled with your OS It is also freely available
for various platforms at http www xfree org
A Common Lisp compiler which supp orts the latest de nition of Common Lisp Steele s
Common Lisp second edition XTAG has b een tested on Lucid Common Lisp SPARC
Solaris V ersion Allegro CL is no longer directly supp orted however there have
b een third party p orts to recentversions of Allegro CL
CLX version or higher CLX is the Lisp equivalent to the Xlib package written in C
Mark Kantrowitz s Lisp Utilities from CMU logical pathnames and defsystem
A patched version of CLX Version for SunOS and the CMU Lisp Utilities are
provided in our ftp directory for your convenience However we ask that you refer to the
appropriate sources for up dates
The morphology database comp onent Karp et al no longer under licensing re
strictions is available as a separate download from the XTAG web page see ab ove for URL
The syntactic database comp onent is also available as part of the XTAG system Egedi
and Martin
CHAPTER OVERVIEW OF THE XTAG SYSTEM
More information can b e obtained on the XTAGweb page at
http www cis upenn edu xtag
Chapter
Underview
The morphology syntactic and tree databases together comprise the English grammar A
lexical item that is not in the databases receives a default tree selection and features for its
part of sp eech and morphology In designing the grammar a decision was made early on to
err on the side of acceptance whenever there are con icting opinions as to whether or not a
construction is grammatical In this sense the XTAG English grammar is intended to function
primarily as an acceptor rather than a generator of English sentences The range of syntactic
phenomena that can be handled is large and includes auxiliaries including inv ersion cop
ula raising and small clause constructions topicalization relative clauses in nitives gerunds
passives adjuncts it clefts wh clefts PRO constructions noun noun mo di cations extrap o
sition determiner sequences genitives negation noun verb contractions clausal adjuncts and
imp eratives
Sub categorization Frames
Elementary trees for non auxiliary verbs are used to represent the linguistic notion of sub cate
gorization frames The anchor of the elementary tree sub categorizes for the other elements that
app ear in the tree forming a clausal or sentential structure Tree families group together trees
b elonging to the same sub categorization frame Consider the following uses of the verb buy
Srini b oughta book
Srini b ought Beth a b o ok
In sentence the verb buy sub categorizes for a direct ob ject NP The elementary tree
anchored by buy is shown in Figure a and includes no des for the NP complementofbuy and
for the NP sub ject In addition to this declarative tree structure the tree family also contains
the trees that would b e related to each other transformationally in a movement based approach
i e passivization imp eratives wh questions relative clauses and so forth Sentence shows
that buy also sub categorizes for a double NP ob ject This means that buy also selects the
double NP ob ject sub categorization frame or tree family with its own set of transformationally
related sentence structures Figure b shows the declarative structure for this set of sentence
structures
CHAPTER UNDERVIEW
Sr Sr
NP0↓ VP NP0↓ VP
V NP1↓ V NP1↓ NP2↓
bought bought
a b
Figure Di erent sub categorization frames for the verb buy
Complements and Adjuncts
Complements and adjuncts havevery di erent structures in the XTAG grammar Complements
are included in the elementary tree anchored bytheverb that selects them while adjuncts do
not originate in the same elementary tree as the verb anchoring the sentence but are instead
added to a structure by adjunction The contrasts b etween complements and adjuncts have b een
extensively discussed in the linguistics literature and the classi cation of a given element as one
or the other remains a matter of debate see Rizzi Larson Jackendo
Larson Cinque Ob ernauer Lasnik and Saito and Chomsky
The guiding rule used in developing the XTAG grammar is whether or not the sentence
is ungrammatical without the questioned structure Consider the following sentences
Srini b oughta book
Srini b ought a b o ok at the b o okstore
Srini arranged for a ride
Srini arranged
Prep ositional phrases frequently occur as adjuncts and when they are used as adjuncts
they have a tree structure such as that shown in Figure a This adjunction tree would
adjoin into the tree shown in Figure a to generate sentence There are verbs however
such as arrange hunger and di erentiate that take prep ositional phrases as complements
Sentences and clearly show that the prep ositional phrase are not optional for arrange
For these sentences the prep ositional phrase will b e an initial tree as shown in Figure b
that substitutes into an elementary tree such as the one anchored by the verb arrange in
Figure c
Virtually all parts of sp eech except for main verbs function as both complements and
adjuncts in the grammar More information is available in this rep ort on various parts of
Iteration of a structure can also b e used as a diagnostic Srini bought a book at the bookstore on Walnut
Street for a friend VPr Sr
PP VP* PP NP0↓ VP
P NP↓ P NP↓ V PP ↓
at for arranged
a b c
Figure Trees illustrating the di erence b etween Complements and Adjuncts
sp eech as complements adjectives e g section nouns e g section and prep ositions
e g section and as adjuncts adjectives section adverbs section nouns
section and prep ositions section
Non S constituents
Although sentential trees are generally considered to be sp ecial cases in any grammar insofar
as they make up a starting category it is the case that any initial tree constitutes a phrasal
constituent These initial trees may have substitution no des that need to be lled by other
initial trees and may b e mo di ed by adjunct trees exactly as the trees ro oted in S Although
grouping is p ossible according to the heads or anchors of these trees we have not found any
classi cation similar to the sub categorization frames for verbs that can be used by a lexical
entry to group select a set of trees These trees are selected one by one by each lexical item
according to each lexical item s idiosyncrasies The grammar describ ed by this technical rep ort
places them into several les for ease of use but these les do not constitute tree families in
the way that the sub categorization frames do
Case Assignment
Approaches to Case
Case in GB theory
GB Government and Binding theory prop oses the following case lter as a requirementon
S structure
Case Filter Every overt NP must b e assigned abstract case Haegeman
There are certain problems with applying the case lter as a requirement at the level of S structure These
issues are not crucial to the discussion of the English XTAG implementation of case and so will not b e discussed
here Interested readers are referred to Lasnik and Uriagereka
CHAPTER UNDERVIEW
Abstract case is taken to be universal Languages with rich morphological case marking
such as Latin and languages with very limited morphological case marking like English are all
presumed to have full systems of abstract case that di er only in the extent of morphological
realization
In GB abstract case is argued to be assigned to NP s by various case assigners namely
verbs prep ositions and INFL Verbs and prep ositions are said to assign accusative casetoNP s
that they govern and INFL assigns nominativecase to NP s that it governs These governing
categories are constrained as to where they can assign case by means of barriers based on
minimality conditions although these are relaxed in exceptional case marking situations
The details of the GB analysis are b eyond the scop e of this technical rep ort but see Chomsky
for the original analysis or Haegeman for an overview Let it su ce for us to say
that the notion of abstract case and the case lter are useful in accounting for a number of
phenomena including the distribution of nominative and accusative case and the distribution
of overt NP s and empty categories suchasPRO
Minimalism and Case
A ma jor conceptual di erence between GB theories and Minimalism is that in Minimalism
lexical items carry their features with them rather than b eing assigned their features based on
the no des that they end up at For nouns this means that they carry case with them and that
their case is checked when they are in SPEC p osition of AGR or AGR which subsequently
s o
disapp ears Chomsky
Case in XTAG
The English XTAG grammar adopts the notion of case and the case lter for many of the same
reasons argued in the GB literature Howev er in some resp ects the English XTAG grammar s
implementation of case more closely resembles the treatment in Chomsky s Minimalism frame
work Chomsky than the system outlined in the GB literature Chomsky As in
Minimalism nouns in the XTAG grammar carry case with them whichiseventually checked
However in the XTAG grammar noun cases are checked against the case values assigned by
the verb during the uni cation of the feature structures Unlike Chomsky s Minimalism there
are no separate AGR no des the case checking comes from the verbs directly Case assignment
from the verb is more like the GB approach than the requirement of a SPEC head relationship
in Minimalism
Most nouns in English do not have separate forms for nominative and accusative case and
so they are ambiguous b etween the two Pronouns of course are morphologically marked for
wtheNP case and each carries the appropriate case in its feature Figures a and b sho
tree anchored bya noun and a pronoun resp ectively along with the feature values asso ciated
with eachword Note that books simply gets the default case nom acc while she restricts the
case to b e nom NP pron : <1> NP wh : <2> - pron : <1> case : <3> nom/acc wh : <2> agr : <4> case : <3> nom/acc agr : <4>
N agr : <4> case : <3> pron : <1> N pron : <1> wh : <2> wh : <2> pron : + case : <3> refl : - agr : <4> case : nom poss : - agr : 3rdsing : - agr : gen : fem num : plur 3rdsing : + pers : 3 num : sing wh : - pers : 3
books she
a b
Figure Lexicalized NP trees with case markings
Case Assigners
Prep ositions
Case is assigned in the XTAG English grammar bytwo lexical categories verbs and prep ositions
Prep ositions assign accusativecase acc through their assign case feature whichislinked
directly to the case feature of their ob jects Figure a shows a lexicalized prep osition
tree while Figure b shows the same tree with the NP tree from Figure a substituted
into the NP p osition Figure c is the tree in Figure b after uni cation has taken place
Note that the case ambiguityof books has b een resolved to accusative case
Verbs
Verbs are the other part of sp eech in the XTAG grammar that can assign case Because XTAG
do es not distinguish INFL and VP no des verbs must provide case assignment on the sub ject
p osition in addition to the case assigned to their NP complements
Assigning case to NP complements is handled by building the case values of the complements
directly into the tree that the case assigner the verb anchors Figures a and b sho w
an S tree that would be anchored by a transitive and ditransitive verb resp ectively Note
that the case assignments for the NP complements are already in the tree even though there
is not yet a lexical item anchoring the tree Since every verb that selects these trees and other
For also assigns case as a complementizer See section for more details
Features not p ertaining to this discussion have b een taken out to improve readability and to make the trees
easier to t onto the page
The diamond marker indicates the anchor s of a structure if the tree has not yet b een lexicalized
CHAPTER UNDERVIEW
PP assign-case : <5> wh : <6>
PP assign-case : <1> acc P assign-case : <5> NP case : <5> wh : <2> - assign-case : acc wh : <6> pron : <1> wh : <2> case : <3> agr : <4> P assign-case : <1> NP agr : <3> 3rdsing : - num : plur pers : 3 PP pron : <4> of N pron : <1> assign-case : <1> case : <1> wh : <2> wh : <2> wh : <2> case : <3> nom/acc agr : <4> agr : 3rdsing : - num : plur of N pron : <4> pers : 3 wh : <2> P assign-case : <1> NP↓ case : <1> wh : - case : <1> assign-case : acc wh : <2> agr : <3>
of books
books
a b c
Figure Assigning case in prep ositional phrases
trees in each resp ective sub categorization frame assigns the same case to the complements
building case features into the tree has exactly the same result as putting the case feature value
in each verb s lexical entry
Sr Sr assign-case : <3> assign-case : <3> agr : <4> agr : <4>
NP0↓ wh : - VP assign-case : <3> NP0↓ wh : - VP assign-case : <3> case : <3> agr : <4> case : <3> agr : <4> agr : <4> assign-case : <1> agr : <4> assign-case : <1> agr : <2> agr : <2>
V◊ assign-case : <1> NP1↓ case : acc V◊ assign-case : <1> NP1↓ case : acc NP2↓ case : acc
agr : <2> agr : <2>
a b
Figure Case assignment to NP arguments
The case assigned to the sub ject p osition varies with verb form Since the XTAG grammar
treats the in ected verb as a single unit rather than dividing it into INFL and V no des case
along with tense and agreement is expressed in the features of verbs and must b e passed in the
appropriate manner The trees in Figure show the path of linkages that joins the assign
case feature of the V to the case feature of the sub ject NP The morphological form of
the verb determines the value of the assign case feature Figures a and b show
the same tree anchored by di erent morphological forms of the verb sing whichgive di erent
values for the assign case feature
Sr assign-case : <3> agr : <4> Sr assign-case : <3> agr : <4>
NP0↓ wh : - VP assign-case : <3> case : <3> agr : <4> agr : <4> assign-case : <1> ↓ agr : <2> NP0 wh : - VP assign-case : <3> case : <3> agr : <4> agr : <4> assign-case : <1> agr : <2> V assign-case : <1> NP1↓ case : acc agr : <2> agr : pers : 3 num : sing V assign-case : <1> NP1↓ case : acc 3rdsing : + agr : <2> assign-case : nom assign-case : none mode : ind mode : ger
sings singing
a b
Figure Assigning case according to verb form
The adjunction of an auxiliary verb onto the VP no de breaks the assign case link from
the main V replacing it with a link from the auxiliary verb instead The progressive form of
the verb in Figure b has the feature value assign case none but this is overridden
by the adjunction of the appropriate form of the auxiliary word be Figure a shows the
lexicalized auxiliary tree while Figure b shows it adjoined into the transitive tree shown
in Figure b The case value passed to the sub ject NP is now nom nominative
PRO in a uni cation based framework
assign case none as the Tensed forms of a verb assign nominative case and untensed forms
progressive form of the verb sing do es in Figure b This is di erent than assigning no case
at all as one form of the in nitive marker to do es See Section for more discussion of this
sp ecial case The distinction of a case none from no case is indicative of a divergence from the
standard GB theory In GB theory the absence of case on an NP means that only PRO can
ll that NP With feature uni cation as is used in the FB LTAG grammar the absence of case
on an NP means that any NP can ll it regardless of its case This is due to the mechanism
of uni cation in which if something is unsp eci ed it can unify with anything Thus we have
Again the feature structures shown have b een restricted to those that p ertain to the V NP interaction
See section for a more complete explanation of how this relinking o ccurs
CHAPTER UNDERVIEW
Sr assign-case : <3> agr : <4>
NP0↓ wh : - VPr assign-case : <3> case : <3> agr : <4> agr : <4> agr : <1> VPr assign-case : <2> agr : <1> assign-case : <2>
V agr : <1> VP assign-case : <2> assign-case : <5> agr : pers : 3 agr : <6> num : sing V agr : <1> VP* 3rdsing : + assign-case : <2> assign-case : nom agr : pers : 3 num : sing is V NP ↓ 3rdsing : + assign-case : <5> 1 case : acc agr : <6> assign-case : nom assign-case : none mode : ger
is singing
a b
Figure Prop er case assignment with auxiliary verbs
a sp eci c case none to handle verb forms that in GB theory do not assign case PRO is the
only NP with case none Note that although we are drawn to this treatment by our use of
uni cation for feature manipulation our treatment is very similar to the assignment of null
case to PRO in Chomsky and Lasnik Watanab e also prop oses a very similar
approach within Chomsky s Minimalist framework
See Sections and for additional discussion of PRO
CHAPTER UNDERVIEW
Part II
Verb Classes
Chapter
Where to Find What
The two page table that follows gives an overview of what typ es of trees occur in various tree
families with p ointers to discussion in this rep ort An entry in a cell of the table indicates that
the tree s for the construction named in the row header are included in the tree family named
in the column header Entries are of two typ es If the particular tree s are displayed and or
discussed in this rep ort the entry gives a page number reference to the relevant discussion or
p
gure Otherwise a indicates inclusion in the tree family but no gure or discussion related
sp eci cally to that tree in this rep ort Blank cells indicate that there are no trees for the
construction named in the row header in the tree family named in the column header Two
tables are given b elow The rst one gives the expansion of abbreviations in the table headers
The second table gives the name given to each tree family in the actual XTAG grammar This
makes it easier to nd the description of each tree family in Chapter and to compare the
description with the online XTAG grammar
Since Chapter has a brief discussion and a declarative tree for every tree family page references are given
only for other sections in which discussion or tree diagrams app ear
CHAPTER WHERE TO FIND WHAT
Abbreviation Full Name
Sent Sub j w to Sentential Sub ject with to PP complement
Pred Mult wd ARB P Predicative Multi word PP with Adv Prep anchors
Pred Mult wd A P Predicative Multi word PP with Adj Prep anchors
Pred Mult wd N P Predicative Multi word PP with Noun Prep anchors
Pred Mult wd P P Predicative Multi word PP with twoPrepanchors
Pred Mult wd no int mo d Predicative Multi word PP with no internal mo di cation
Pred Sent Sub j ARB P Predicative PP with Sentential Sub ject and Adv Prep anchors
Pred Sent Sub j A P Predicative PP with Sentential Sub ject and Adj Prep anchors
Pred Sent Sub j Conj P Predicative PP with Sentential Sub ject and Conj Prep anchors
Pred Sent Sub j N P Predicative PP with Sentential Sub ject and Noun Prep anchors
Pred Sent Sub j P P Predicative PP with Sentential Sub ject and two Prep anchors
t Sub j no int mo d Predicative PP with Sentential Sub ject no internal mo di cation Pred Sen
Pred Lo cative Predicativeanchored by a Lo cativeAdverb
Pred A Sent Sub j Comp Predicative AdjectivewithSentential Sub ject and Complement
Sentential Comp with NP Sentential Complement with NP
Pred Mult wd V P Predicative Multi word with Verb Prep anchors
Adj Sm Cl w Sentential Sub j Adjective Small Clause with Sentential Sub ject
NP Sm Clause w Sentential Sub j NP Small Clause with Sentential Sub ject
PP Sm Clause w Sentential Sub j PP Small Clause with Sentential Sub ject
NP Sm Cl w Sent Comp NP Small Clause with Sentential Complement
Adj Sm Cl w Sent Comp Adjective Small Clause with Sentential Complement
ExhaustivePPSm Cl Exhaustive PP Small Clause
Ditrans LightVerbs w PP Shift DitransitiveLightVerbs with PP Shift
Ditrans LightVerbs w o PP Shift DitransitiveLightVerbs without PP Shift
Y N question Yes No question
Wh mov NP complement Wh moved NP complement
Wh mov S comp Wh moved S complement
Wh mov Adj comp Wh moved Adjective complement
Wh mov ob ject of a P Wh moved ob ject of a P
Wh mov PP Wh moved PP
Topic NP complement Topicalized NP complement
Det gerund Determiner gerund
Rel cl on NP comp Relative clause on NP complement
Rel cl on PP comp Relative clause on PP complement
Rel cl on NP ob ject of P Relative clause on NP ob ject of P
Pass with wh moved sub j Passive with wh moved sub ject with and without by phrase
Pass w wh mov ind ob j Passive with wh mov ed indirect ob ject with and without by phrase
Pass w wh mov ob j of the by phrase Passive with wh moved ob ject of the by phrase
Pass w wh mov by phrase Passive with wh moved by phrase
Trans Idiom with V D and N Transitive Idiom with Verb Det and Noun anchors
Idiom with V D N IdiomwithV D andNanchors
Idiom with V D A N IdiomwithV D A andNanchors
Idiom with V N Idiom with V and N anchor
Idiom with V A N Idiom with V A and N anchors
Idiom with V D N P Idiom with V D N and Prep anchors
Idiom with V D A N P Idiom with V D A N and Prep anchors
Idiom with V N P Idiom with V N and Prep anchors
Idiom with V A N P IdiomwithV A N andPrepanchors
Full Name XTAGName
IntransitiveSentential Sub ject Ts V
Sentential Sub ject with to complement Ts Vtonx
PP Small Clause with Adv and Prep anchors Tnx ARBPnx
PP Small Clause with Adj and Prep anchors Tnx APnx
PP Small Clause with Noun and Prep anchors Tnx NPnx
PP Small Clause with Prep anchors Tnx PPnx
PP Small Clause with Prep and Noun anchors Tnx PNaPnx
PP Small Clause with Sentential Sub ject and Adv and Prep anchors Ts ARBPnx
PP Small Clause with Sentential Sub ject and Adj and Prep anchors Ts APnx
PP Small Clause with Sentential Sub ject and Noun and Prep anchors Ts NPnx
PP Small Clause with Sentential Sub ject and Prep anchors Ts PPnx
PP Small Clause with Sentential Sub ject and Prep and Noun anchors Ts PNaPnx
Exceptional Case Marking TXnx Vs
Lo cative Small Clause with Ad anchor Tnx nx ARB
Predicative AdjectivewithSentential Sub ject and Complement Ts A s
Transitive Tnx Vnx
Ditransitive with PP shift Tnx Vnx tonx
Ditransitive Tnx Vnx nx
Ditransitive with PP Tnx Vnx pnx
Sentential Complement with NP Tnx Vnx s
IntransitiveVerb Particle Tnx Vpl
TransitiveVerb Particle Tnx Vplnx
DitransitiveVerb Particle Tnx Vplnx nx
Intransitiv e with PP Tnx Vpnx
Sentential Complement Tnx Vs
LightVerbs Tnx lVN
Ditransitive LightVerbs with PP Shift Tnx lVN Pnx
Adjective Small Clause with Sentential Sub ject Ts Ax
NP Small Clause with Sentential Sub ject Ts N
PP Small Clause with Sentential Sub ject Ts Pnx
Predicative Multi word with Verb Prep anchors Tnx VPnx
Adverb It Cleft TItVad s
NP It Cleft TItVnx s
PP It Cleft TItVpnx s
Adjective Small Clause Tree Tnx Ax
Adjective Small Clause with Sentential Complement Tnx A s
Equative BE Tnx BEnx
NP Small Clause Tnx N
NP with Sentential Complement Small Clause Tnx N s
PP Small Clause Tnx Pnx
Exhaustive PP Small Clause Tnx Px
Intransitive Tnx V
Intransitive with Adjective Tnx Vax
TransitiveSentential Sub ject Ts Vnx
Idiom with V D and N Tnx VDN
Idiom with V D A and N anchors Tnx VDAN
Idiom with V and N anchors Tnx VN
Idiom with V A and N anchors Tnx VAN
Idiom with V D N and Prep anchors Tnx VDN Pnx
Idiom with V D A N and Prep anchors Tnx VDAN Pnx
Idiom with V N and Prep anchors Tnx VN Pnx
Idiom with V A N and Prep anchors Tnx VAN Pnx
CHAPTER WHERE TO FIND WHAT
Tree families mo d t mo d t P tial Sub j P A P P no in ARB P N P ten to Sub j Comp e Sub j ARB P Sub j A P Sub j N P Sub j P Sub j no in t w e Sen t t t t t Sen Mult wd Mult wd Mult wd Mult wd Mult wd Sen Sen Sen Sen Lo cativ A Sen Sub j t
transitiv Constructions
In Sen Pred Pred Pred Pred Pred Pred Pred Pred Pred Pred ECM Pred Pred
p p p p p p p p p p p p
Declarative
Passive w w o by phrase
Y N quest
p p p p p p p p p p p p p p p
Wh moved sub ject
Wh mov NP complement DO or IO
Wh mov S comp
Wh mov Adj or Adv comp
p p p p p
Wh mov ob ject of a P
p p p
Wh mov PP
Topic NP comp
p p p p p
Imp erative
Det gerund
p p p p
NP gerund
Ergative
p p p p p p p
Rel cl on sub j w NP
p p p p p p p
Rel cl on sub j w Comp
Rel cl on NP comp DO IO w NP
Rel cl on NP comp DO IO w
Comp
p p
Rel cl on PP comp w pied piping
p p p p p p p p p p
Rel cl on NP ob ject of P w NP
p p p p p p p p p
Rel cl on NP ob ject of P w Comp
p p p p p p p p p p p p p p
Rel cl on adjunct w PP
p p p p p p p p p p p p p p
Rel cl on adjunct w Comp
Pass w wh mov sub j
Pass w wh mov ind ob j
Pass w wh mov ob j of by phrase
Pass w wh mov by phrase
Tree families tial Sub j t tial Sub j tial Sub j ten article article ten ten article V P erb P erb P tVs erb P tVs e with PP shift e ewithPP eV eV ewithPP e eV Ligh tial Comp with NP tial Complemen ten ten transitiv transitiv
ransitiv ransitiv rans Constructions
T Ditransitiv Ditransitiv Ditransitiv Sen In T Ditransitiv In Sen T Ditrans Ligh Adj Sm Cl w Sen NP Sm Cl w Sen PP Sm Cl w Sen Pred Mult wd
p p p p p p p p p p p
Declarative
p p p p p p p p
Passive w w o by phrase
Y N quest
p p p p p p p p p p p p p p p
Wh moved sub ject
p p p p p
Wh mov NP complement DO or IO
p p
Wh mov S comp
p
Wh mov Adj or Adv comp
p p p p
Wh mov ob ject of a P
p p p
Wh mov PP
p p p p p p p
Topic NP comp
p p p p p p p p p p p p
Imp erative
p p p p p p p p p p p p
Det gerund
p p p p p p p p p p p p
NP gerund
Ergative
p p p p p p p p p p p p p
Rel cl on sub j w NP
p p p p p p p p p p p p p
Rel cl on sub j w Comp
p p p p p p p p p
Rel cl on NP comp DO IO w NP
p p p p p p p p p
Rel cl on NP comp DO IO w
Comp
p p p p p p p p p
Rel cl on PP comp w pied piping
p p p p p p p p p p
Rel cl on NP ob ject of P w NP
p p p p p p p p p p
Rel cl on NP ob ject of P w Comp
p p p p p p p p p p p p p p p p
Rel cl on adjunct w PP
p p p p p p p p p p p p p p p p
Rel cl on adjunct w Comp
p p
Parenthetical quoting clause
p p
Past participal as arg Adj
p p
Past participial NP pre mo d
p p p p p p p p p
Pass w wh mov sub j
p p p p p p
Pass w wh mov ind ob j
p p p p p p p p p
Pass w wh mov ob j of by phrase
p p p p p p p p
Pass w wh mov by phrase
CHAPTER WHERE TO FIND WHAT
Tree families e Cl t Comp t Comp tial Sub j ten e e with Adjectiv ePPSm BE eSen e Small Clause erb It Cleft transitiv transitiv
ransitiv Constructions
Adv NP It Cleft PP It Cleft Adj Adj Sm Cl w Sen Equativ NP Small Clause NP Sm Cl w Sen PP Small Clause Exhaustiv In In T
p p p p p p p
Declarative
Passivew w oby phrase
p p p
Y N quest
p p p p p p
Wh moved sub ject
p p p
Wh mov NP complement DO or IO
Wh mov S comp
p p
Wh mov Adj or Adv comp
p
Wh mov ob ject of a P
p p
Wh mov PP
p p
Topic NP comp
p p p p p p p p
Imp erative
Det gerund
p p p p p p p p
NP gerund
Ergative
p p p p p p p p
Rel cl on sub j w NP
p p p p p p p p
Rel cl on sub j w Comp
Rel cl on NP comp DO IO w NP
Rel cl on NP comp DO IO w
Comp
p
Rel cl on PP comp w pied piping
p
Rel cl on NP ob ject of P w NP
p
Rel cl on NP ob ject of P w Comp
p p p p p p p p p p p p
Rel cl on adjunct w PP
p p p p p p p p p p p p
Rel cl on adjunct w Comp
p
Participial NP pre mo d
Pass w wh mov sub j
Pass w wh mov ind ob j
Pass w wh mov ob j of by phrase
Pass w wh mov by phrase
Tree families Constructions
Idiom with V D N Idiom with V D A N Idiom with V N Idiom with V A N Idiom with V D N P Idiom with V D A N Idiom P with V N P Idiom with V A N P
p p p p p p p p
Declarative
p p p p p p p p
Passivew w oby phrase
Y N quest
p p p p p p p p
Wh moved sub ject
Wh mov NP complement DO or IO
Wh mov S comp
Wh mov Adj or Adv comp
Wh mov ob ject of a P
Wh mov PP
Topic NP comp
p p p p p p p p
Imp erative
Det gerund
p p p p p p p p
NP gerund
Ergative
p p p p p p p p
Rel cl on sub j w NP
p p p p p p p p
Rel cl on sub j w Comp
Rel cl on NP comp DO IO w NP
Rel cl on NP comp DO IO w
Comp
Rel cl on PP comp w pied piping
Rel cl on NP ob ject of P w NP
Rel cl on NP ob ject of P w Comp
p p p p p p p p
Rel cl on adjunct w PP
p p p p p p p p
Rel cl on adjunct w Comp
Pass w wh mov sub j
Pass w wh mov ind ob j
p p p p p p p p
Pass w wh mov ob j of by phrase
p p p p p p p p
Pass w wh mov by phrase
p p p p
Outer Pass w and wo by phrase
p p p p p
Outer Pass w Rel cl on sub j w
Comp
p p p p p
Outer Pass w Rel cl on sub j w
NP
Chapter
Verb Classes
Each main verb in the syntactic lexicon selects at least one tree family sub categorization
frame Since the tree database and syntactic lexicon are already separated for space e ciency
see Chapter each verb can e ciently select a large number of trees by sp ecifying a tree
family as opp osed to each of the individual trees This approach allows for a considerable
reduction in the numb er of trees that must b e sp eci ed for anygiven verb or form of a verb
There are currently tree families in the system This chapter gives a brief description
of each tree family and shows the corresp onding declarative tree along with any p eculiar
characteristics or trees It also indicates which transformations are in each tree family and
gives the number of verbs that select that family A few sample verbs are given along with
example sentences
Intransitive Tnx V
Description This tree family is selected byverbs that do not require an ob ject complement
of any typ e Adverbs prep ositional phrases and other adjuncts may adjoin on but are
not required for the sentences to b e grammatical verbs select this family
dance Examples eat sleep
Alate
Seth slept
Hyun danced
Declarative tree See Figure
Other available trees wh moved sub ject sub ject relative clause with and without comp
adjunct gap less relative clause with comp adjunct gap less relative clause with PP
pied piping imp erative determiner gerund NP gerund pre nominal participal
Auxiliary verbs are handled under a di erentmechanism See Chapter for details
See section for explanation of tree families
An explanation of the naming convention used in naming the trees and tree families is available in Ap
p endix D
Before lexicalization the indicates the anchor of the tree
Numbers given are as of August and are sub ject to some change with further development of the
grammar Sr
NP0↓ VP
V◊
Figure DeclarativeIntransitiveTree nx V
Transitive Tnx Vnx
Description This tree family is selected byverbs that require only an NP ob ject complement
The NP s may b e complex structures including gerund NP s and NP s that take sentential
complements This do es not include lightverb constructions see sections and
verbs select the transitive tree family
Examples eat dance take like
Alate an apple
Seth danced the tango
Hyun is taking an algorithms course
Anoop likes the fact that the semester is nished
Declarative tree See Figure
Sr
NP0↓ VP
V◊ NP1↓
Figure DeclarativeTransitiveTree nx Vnx
Other available trees wh moved sub ject wh moved ob ject sub ject relative clause with and
without comp adjunct gap less relativeclause with comp with PP pied piping ob ject
relative clause with and without comp imp erative determiner gerund NP gerund passive
with by phrase passivewithoutby phrase passive with wh moved sub ject and by phrase
passive with wh moved sub ject and no by phrase passive with wh moved ob ject out of the
by phrase passive with wh moved by phrase passive with relative clause on sub ject and
by phrase with and without comp passive with relative clause on sub ject and no by phrase
with and without comp passive with relative clause on ob ject on the by phrase with and
without comp with PP pied piping gerund passivewithby phrase gerund passive without
by phrase ergative ergative with wh moved sub ject ergative with sub ject relative clause
with and without comp ergative with adjunct gap less relative clause with comp with
CHAPTER VERB CLASSES
PP pied piping In addition two other trees that allow transitive verbs to function as
adjectives e g the stopped truck are also in the family
Ditransitive Tnx Vnx nx
Description This tree family is selected byverbs that take exactly two NP complements It
do es not include verbs that undergo the ditransitive verb shift see section The
apparent ditransitive alternates involving verbs in this class and b enefactive PP s e g
John baked a cake for Mary are analyzed as transitives see section with a PP
adjunct Benefactiv es are taken to b e adjunct PP s b ecause they are optional e g John
bakedacake vs John bakedacake for Mary verbs select the ditransitive tree family
Examples ask cook win
Christy asked Mike a question
Doug cooked his father dinner
Dania won her sister a stu ed animal
Declarative tree See Figure
Sr
NP0↓ VP
V◊ NP1↓ NP2↓
Figure Declarative DitransitiveTree nx Vnx nx
Other available trees wh moved sub ject wh moved direct ob ject wh moved indirect ob
ject sub ject relative clause with and without comp adjunct gap less relative clause
with comp with PP pied piping direct ob ject relative clause with and without comp in
direct ob ject relative clause with and without comp imp erative determiner gerund NP
gerund passive with by phrase passive without by phrase passive with wh moved sub ject
and by phrase passive with wh moved sub ject and no by phrase passive with wh moved
ob ject out of the by phrase passive with wh moved by phrase passive with wh moved
indirect ob ject and by phrase passive with wh moved indirect ob ject and no by phrase
passive with relative clause on sub ject and by phrase with and without comp passivewith
relative clause on sub ject and no by phrase with and without comp passive with relative
clause on ob ject of the by phrase with and without comp with PP pied piping passive
with relative clause on the indirect ob ject and by phrase with and without comp passive
with relative clause on the indirect ob ject and no by phrase with and without comp pas
sive with without by phrase with adjunct gap less relative clause with comp with PP
pied piping gerund passive with by phrase gerund passive without by phrase
Ditransitive with PP Tnx Vnx pnx
Description This tree family is selected by ditransitiveverbs that take a noun phrase followed
by a prep ositional phrase The prep osition is not constrained in the syntactic lexicon The
prep osition must be required and not optional that is the sentence must be ungram
matical with just the noun phrase e g John put the table No verbs therefore should
select b oth this tree family and the transitive tree family see section This tree
family is also distinguished from the ditransitive verbs such as give that undergo verb
shifting see section There are verbs that select this tree family
Examples associate put refer
Rostenkowski associated money with power
He put his reputation on the line
He referred al l questions to his attorney
Declarative tree See Figure
Sr
NP0↓ VP
V◊ NP1↓ PP2↓
Figure Declarative Ditransitive with PP Tree nx Vnx pnx
Other available trees wh moved sub ject wh moved direct ob ject wh moved ob ject of PP
wh moved PP sub ject relative clause with and without comp adjunct gap less relative
clause with comp with PP pied piping direct ob ject relative clause with and without
comp ob ject of PP relative clause with and without comp with PP pied piping imp er
ative determiner gerund NP gerund passive with by phrase passive without by phrase
passive with wh moved sub ject and by phrase passive with wh moved sub ject and no by
phrase passive with wh moved ob ject out of the by phrase passive with wh moved by
phrase passive with wh moved ob ject out of the PP and by phrase passive with wh moved
ob ject out of the PP and no by phrase passiv e with wh moved PP and by phrase passive
with wh moved PP and no by phrase passive with relative clause on sub ject and by phrase
with and without comp passive with relative clause on sub ject and no by phrase with and
without comp passive with relative clause on ob ject of the by phrase with and without
comp with PP pied piping passive with relative clause on the ob ject of the PP and by
phrase with and without comp with PP pied piping passive with relative clause on the
ob ject of the PP and no by phrase with and without comp with PP pied piping passive
with and without by phrase with adjunct gap less relative clause with comp with PP
pied piping gerund passive with by phrase gerund passive without by phrase
CHAPTER VERB CLASSES
Ditransitive with PP shift Tnx Vnx tonx
Description This tree family is selected by ditransitive verbs that undergo a shift to a to
prep ositional phrase These ditransitiveverbs are clearly constrained so that when they
shift the prep ositional phrase must start with to This is in contrast to the Ditransitives
with PP in section in whichverbs may app ear in NP V NP PP constructions with
a variety of prep ositions Both the dative shifted and non shifted PP complement trees
are included verbs select this family
Examples give promise tel l
Bil l gave Hil lary owers
Bil l gave owers to Hil lary
Whitman promised the voters a tax cut
Whitman promised a tax cut to the voters
Pinnochino told Gepetto a lie
Pinnochino told a lie to Gepetto
Declarative tree See Figure
Sr
NP0↓ VP
Sr
V◊ NP1↓ PP2
NP ↓ VP P2 NP2↓ 0
to V◊ NP2↓ NP1↓
a b
Figure Declarative Ditransitive with PP shift Trees nx Vnx Pnx a and
nx Vnx nx b
Other available trees Non shifted wh moved sub ject wh moved direct ob ject wh moved
indirect ob ject sub ject relative clause with and without comp adjunct gap less relative
clause with comp with PP pied piping direct ob ject relative clause with comp with PP
pied piping indirect ob ject relative clause with and without comp with PP pied piping
imp erative NP gerund passive with by phrase passive without by phrase passive with
wh moved sub ject and by phrase passive with wh moved sub ject and no by phrase passive
with wh moved ob ject out of the by phrase passive with wh moved by phrase passive
with wh moved indirect ob ject and by phrase passive with wh moved indirect ob ject and
with relative clause on sub ject and by phrase with and without no by phrase passive
comp passive with relative clause on sub ject and no by phrase with and without comp
passive with relative clause on ob ject of the by phrase with and without comp with PP
pied piping passive with relative clause on the indirect ob ject and by phrase with and
without comp with PP pied piping passive with relative clause on the indirect ob ject
and no by phrase with and without comp with PP pied piping passive with without
by phrase with adjunct gap less relative clause with comp with PP pied piping gerund
passivewithby phrase gerund passive without by phrase
Shifted wh moved sub ject wh moved direct ob ject wh moved ob ject of PP wh moved
PP sub ject relative clause with and without comp adjunct gap less relative clause with
comp with PP pied piping direct ob ject relative clause with comp with PP pied piping
ob ject of PP relative clause with and without comp with PP pied piping imp erative
determiner gerund NP gerund passivewithby phrase passive without by phrase passive
with wh moved sub ject and by phrase passive with wh moved sub ject and no by phrase
passive with wh moved ob ject out of the by phrase passive with wh moved by phrase
passive with wh moved ob ject out of the PP and by phrase passive with wh moved ob ject
outofthePPandnoby phrase passive with wh moved PP and by phrase passivewith
wh moved PP and no by phrase passive with relative clause on sub ject and by phrase
with and without comp passive with relative clause on sub ject and no by phrase with and
without comp passive with relative clause on ob ject of the by phrase with and without
comp with PP pied piping passive with relative clause on the ob ject of the PP and by
phrase with and without comp with PP pied piping passive with relative clause on the
ob ject of the PP and no by phrase with and without comp with PP pied piping passive
with without by phrase with adjunct gap less relative clause with comp with PP pied
piping gerund passivewithby phrase gerund passive without by phrase
Sentential Complement with NP Tnx Vnx s
Description This tree family is selected by verbs that take both an NP and a sentential
complement The sentential complement may be in nitive or indicative The typ e of
clause is sp eci ed by each individual verb in its syntactic lexicon entry A given verb
may select more than one typ e of sen tential complement The declarative tree and many
other trees in this family are auxiliary trees as opp osed to the more common initial trees
These auxiliary trees adjoin onto an S no de in an existing tree of the typ e sp eci ed by
the sentential complement This is the mechanism by whichTAGs are able to maintain
long distance dep endencies see Chapter even over multiple emb eddings e g What
did Bil l tel l Mary that John said verbs select this tree family
Examples beg expect tel l
Srini begged Mark to increase his disk quota
his turn Beth told Jim that it was
Declarative tree See Figure
Other available trees wh moved sub ject wh moved ob ject wh moved sentential comple
ment sub ject relative clause with and without comp adjunct gap less relative clause
with comp with PP pied piping ob ject relative clause with and without comp imp erative
determiner gerund NP gerund passivewithby phrase b efore sentential complement pas
sivewithby phrase after sentential complement passive without by phrase passivewith
CHAPTER VERB CLASSES
Sr
NP0↓ VP
V◊ NP1↓ S2*
Figure DeclarativeSentential ComplementwithNPTree nx Vnx s
wh moved sub ject and by phrase b efore sentential complement passive with wh moved
sub ject and by phrase after sentential complement passive with wh moved sub ject and
no by phrase passive with wh moved ob ject out of the by phrase passive with wh moved
by phrase passive with relative clause on sub ject and by phrase b efore sentential comple
ment with and without comp passive with relative clause on sub ject and by phrase after
sentential complement with and without comp passive with relative clause on sub ject and
no by phrase with and without comp passive with without by phrase with adjunct gap
less relative clause with comp with PP pied piping gerund passive with by phrase b efor
sentential complement gerund passive with by phrase after the sentential complement
gerund passivewithoutby phrase parenthetical rep orting clause
Intransitive Verb Particle Tnx Vpl
Description The trees in this tree family are anchored by b oth the verb and the verb parti
cle Both app ear in the syntactic lexicon and together select this tree family Intransitive
verb particles can b e di cult to distinguish from intransitiveverbs with adverbs adjoined
on The main diagnostics for including verbs in this class are whether the meaning is
comp ositional or not and whether there is a transitive version of the verb verb particle
combination with the same or similar meaning The existence of an alternate comp osi
tional meaning is a strong indication for a separate verb particle construction There are
verb verb particle combinations
Examples add up come out sign o
never quite added up The numbers
John nal ly came out of the closet
I think that I wil l sign o now
Declarative tree See Figure
Other available trees wh moved sub ject sub ject relative clause with and without comp
adjunct gap less relative clause with comp with PP pied piping imp erative determiner
gerund NP gerund Sr
NP0↓ VP
V◊ PL◊
Figure DeclarativeIntransitiveVerb Particle Tree nx Vpl
Transitive Verb Particle Tnx Vplnx
Description Verb verb particle combinations that take an NP complement select this tree
family Both the verb and the verb particle are anchors of the trees Particle movement has
b een taken as the diagnostic to distinguish verb particle constructions from intransitives
with adjoined PP s If the alleged particle is able to undergo particle movement in other
words app ear both before and after the direct ob ject then it is judged to be a particle
Items that do not undergo particle movement are taken to be prep ositions In many
but not all of the verb particle cases there is also an alternate prep ositional meaning in
which the lexical item did not mov e e g He looked up the number in the phonebook
He looked the number up Srini looked up the road for Purnima s car He looked the
road up There are verb verb particle combinations
Examples blow o make up pick out
He blew o his linguistics class for the third time
He blew his linguistics class o for the third time
The dyslexic leprechaun made up the syntactic lexicon
The dyslexic leprechaun made the syntactic lexicon up
I would like to pick out a new computer
I would like to pick a new computer out
Declarative tree See Figure
Sr Sr
NP0↓ VP NP0↓ VP
V◊ PL◊ NP1↓ V◊ NP1↓ PL◊
a b
Figure DeclarativeTransitiveVerb Particle Tree nx Vplnx a and nx Vnx pl b
Other available trees wh moved sub ject with particle b efore the NP wh moved sub ject
with particle after the NP wh moved ob ject sub ject relative clause with particle b efore
CHAPTER VERB CLASSES
the NP with and without comp sub ject relative clause with particle after the NP with and
without comp ob ject relative clause with and without comp adjunct gap less relative
clause with particle b efore the NP with comp with PP pied piping adjunct gap less
relative clause with particle after the NP with comp with PP pied piping imp erative
with particle before the NP imp erative with particle after the NP determiner gerund
with particle b efore the NP NP gerund with particle before the NP NP gerund with
particle after the NP passive with by phrase passivewithoutby phrase passive with wh
moved sub ject and by phrase passive with wh moved sub ject and no by phrase passive
with wh moved ob ject out of the by phrase passive with wh moved by phrase passive
with relative clause on sub ject and by phrase with and without comp passive with relative
clause on sub ject and no by phrase with and without comp passive with relative clause on
ob ject of the by phrase with and without comp with PP pied piping passive with without
by phrase with adjunct gap less relative clause with comp with PP pied piping gerund
passivewithby phrase gerund passive without by phrase
Ditransitive Verb Particle Tnx Vplnx nx
Description Verb verb particle combinations that select this tree family take NP comple
ments Both the verb and the verb particle anchor the trees and the verb particle can
o ccur b efore between or after the noun phrases Perhaps b ecause of the complexity of
the sen tence these verbs do not seem to have passive alternations A new bank account
was opened up Michel le by me There are verb verb particle combinations that select
this tree family The exhaustive list is given in the examples
Examples dish out open up pay o rustle up
I opened up Michel le a new bank account
I opened Michel le up a new bank account
I opened Michel le a new bank account up
Declarative tree See Figure
Sr Sr Sr
NP0↓ VP NP0↓ VP NP0↓ VP
V◊ PL◊ NP1↓ NP2↓ V◊ NP1↓ PL◊ NP2↓ V◊ NP1↓ NP2↓ PL◊
a b c
Figure Declarative Ditransitive Verb Particle Tree nx Vplnx nx a
nx Vnx plnx b and nx Vnx nx pl c
Other available trees wh moved sub ject with particle b efore the NP s wh moved sub ject
with particle b etween the NP s wh moved sub ject with particle after the NP s wh moved
indirect ob ject with particle b efore the NP s wh moved indirect ob ject with particle after
the NP s wh moved direct ob ject with particle b efore the NP s wh moved direct ob ject
with particle between the NP s sub ject relative clause with particle b efore the NP s
with and without comp sub ject relative clause with particle between the NP s with
and without comp sub ject relative clause with particle after the NP s with and without
comp indirect ob ject relative clause with particle b efore the NP s with and without comp
indirect ob ject relative clause with particle after the NP s with and without comp direct
ob ject relative clause with particle b efore the NP s with and without comp direct ob ject
relative clause with particle between the NP s with and without comp adjunct gap
less relative clause with comp with PP pied piping imp erative with particle b efore the
NP s imp erative with particle b etween the NP s imp erative with particle after the NP s
determiner gerund with particle b efore the NP s NP gerund with particle b efore the NP s
NP gerund with particle b etween the NP s NP gerund with particle after the NP s
Intransitive with PP Tnx Vpnx
Description The verbs that select this tree family are not strictly intransitive in that they
must b e followed by a prep ositional phrase Verbs that are intransitive and simply can
be followed by a prep ositional phrase do not select this family but instead have the
PP adjoin onto the intransitivesentence Accordingly there should b e no verbs in b oth
this class and the intransitive tree family see section The prep ositional phrase is
not restricted to b eing headed by any particular lexical item Note that these are not
transitive verb particles see section since the head of the PP do es not move
verbs select this tree family
Examples grab impinge provide
Seth grabbed for the brass ring
The noise gradual ly impinged on Dania s thoughts
Agood host provides for everyone s needs
Declarative tree See Figure
Sr
NP0↓ VP
V◊ PP1↓
Figure DeclarativeIntransitive with PP Tree nx Vpnx
Other available trees wh moved sub ject wh moved ob ject of the PP wh moved PP sub ject
relative clause with and without comp adjunct gap less relative clause with comp with
CHAPTER VERB CLASSES
PP pied piping ob ject of the PP relative clause with and without comp with PP pied
piping imp erative determiner gerund NP gerund passivewithby phrase passive without
by phrase passive with wh moved sub ject and by phrase passive with wh moved sub ject
and no by phrase passivewithwh moved by phrase passive with relative clause on sub ject
and by phrase with and without comp passive with relative clause on sub ject and no by
phrase with and without comp passive with relative clause on ob ject of the by phrase with
and without comp with PP pied piping passive with without by phrase with adjunct
gap less relative clause with comp with PP pied piping gerund passive with by phrase
gerund passivewithoutby phrase
Predicative Multi word with Verb Prep anchors Tnx VPnx
Description This tree family is selected by multiple anchor verb prep osition pairs which
together ha ve a non comp ositional interpretation For example think of has the non
comp ositional interpretion involving the inception of a notion or mental entity in addition
to the interpretion in which the agent is thinking ab out someone or something Anchors
for this tree must be able to take both gerunds and regular NP s in the second noun
p osition To allow adverbs to app ear between the verb and the prep osition the trees
contain an extra VP level Several of the verbs which select the Tnx Vpnx family but
which should not have quite the freedom it allows will b e moving to this family for the
next release verb prep osition pairs select this tree family
Examples think of believe in depend on
Calvin thought of a new idea
Hobbes believes in sleeping al l day
Bil l depends on drinking co ee for stimulation
Declarative tree See Figure
S r
NP 0↓ VP1
VP2 PP
V◊ P◊ NP 1↓
Figure Declarative PP ComplementTree nx VPnx
Other available trees wh moved sub ject wh moved ob ject sub ject relative clause with and
without comp adjunct gap less relativeclause with comp with PP pied piping ob ject
relative clause with and without comp imp erative determiner gerund NP gerund passive
with by phrase passivewithoutby phrase passive with wh moved sub ject and by phrase
passive with wh moved sub ject and no by phrase passive with wh moved ob ject out of the
by phrase passive with wh moved by phrase passive with relative clause on sub ject and by
phrase with and without comp passive with relative clause on sub ject and no by phrase
with and without comp passive with relative clause on ob ject on the by phrase with
and without comp with PP pied piping passive with without by phrase with adjunct
gap less relative clause with comp with PP pied piping gerund passive with by phrase
gerund passive without by phrase In addition two other trees that allow transitiveverbs
to function as adjectives e g the thought of idea are also in the family
Sentential Complement Tnx Vs
tential complement The Description This tree family is selected byverbs that take just a sen
sentential complementmaybeoftyp e in nitive indicative or small clause see Chapter
The typ e of clause is sp eci ed byeach individual verb in its syntactic lexicon entry anda
given verb may select more than one typ e of sentential complement The declarative tree
and many other trees in this family are auxiliary trees as opp osed to the more common
initial trees These auxiliary trees adjoin onto an S no de in an existing tree of the typ e
sp eci ed by the sentential complement This is the mechanism by which TAGs are able
to maintain long distance dep endencies see Chapter even over multiple emb eddings
e g What did Bil l think that John said verbs select this tree family
Examples consider think
Dania considered the algorithm unworkable
Srini thought that the program was working
Declarative tree See Figure
Sr
NP0↓ VP
V◊ S1*
Figure Declarative Sentential ComplementTree nx Vs
Other available trees wh moved sub ject wh moved sentential complement sub ject relative
clause with and without comp adjunct gap less relative clause with comp with PP pied
piping imp erative determiner gerund NP gerund parenthetical rep orting clause
Intransitive with Adjective Tnx Vax
Description The verbs that select this tree family take an adjective as a complement The
CHAPTER VERB CLASSES
adjective may be regular comparative or sup erlative It may also be formed from the
sp ecial class of adjectives derived from the transitive verbs e g agitated broken See
section UnliketheIntransitive with PP verbs see section some of these verbs
may also o ccur as bare intransitives as well This distinction is drawn b ecause adjectives
do not normally adjoin onto sentences as prep ositional phrases do Other intransitive
verbs can only o ccur with the adjective and these select only this family The verb class is
also distinguished from the adjective small clauses see section b ecause these verbs
are not raising verbs verbs select this tree family
Examples become grow smel l
The greenhouse became hotter
The plants grew tal l and strong
The owers smel led wonderful
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ AP1 ↓
Figure DeclarativeIntransitive with AdjectiveTree nx Vax
Other available trees wh moved sub ject wh moved adjective how sub ject relative clause
with and without comp adjunct gap less relative clause with comp with PP pied piping
imp erative NP gerund
Transitive Sentential Sub ject Ts Vnx
Description The verbs that select this tree family all take sentential sub jects and are often
referred to as psych verbs since they all refer to some psychological state of mind The
sentential sub ject can b e indicative complementizer required or in nitive complemen
tizer optional verbs that select this tree family
Examples delight impress surprise
that the tea had rosehips in it delighted Christy
to even attempt a marathon impressed Dania
For Jim to have walked the dogs surprise d Beth
Declarative tree See Figure
Other available trees wh moved sub ject wh moved ob ject sub ject relative clause with and
without comp adjunct gap less relative clause with comp with PP pied piping Sr
S0↓ VP
V◊ NP1↓
Figure Declarative Sentential Sub ject Tree s Vnx
Light Verbs Tnx lVN
Description The verb noun pairs that select this tree families are pairs in whichtheinterpre
tation is non comp ositional and the noun contributes argument structure to the predicate
e g The man took a walk vs The man took a radio The verb and the noun o ccur
together in the syntactic database and b oth anchor the trees The verbs in the lightverb
constructions are do give have make and take The noun following the lightverb is usu
ally in a bare in nitive form have a good cry and usually o ccurs with a n However
we include deverbal nominals take a bath give a demonstr ation as well Constructions
with nouns that do not contribute an argument structure have a cigarette give NP a
black eye are excluded In addition to semantic considerations of lightverbs they di er
syntactically from Transitiveverbs section as well in that the noun in the lightverb
construction do es not extract Some of the verb noun anchors for this family like take
aim and take hold disallow determiners while others require particular determiners For
example have think must b e inde nite and singular as attested by the ungrammaticality
of John had the think some thinks Another anchor take leave can occur either bare
or with a possesive pronoun e g John took his leave but not John took the leave
This is accomplished through feature sp eci cation on the lexical entries There are
verb noun pairs that select the lightverb tree
Examples give groan have discussion make comment
The audience gave a col lective groan
We had a big discussion about closing the libraries
The professors made comments on the paper
Declarative tree See Figure
Other available trees wh moved sub ject sub ject relative clause with and without comp
e determiner adjunct gap less relative clause with comp with PP pied piping imp erativ
gerund NP gerund
Ditransitive Light Verbs with PP Shift Tnx lVN Pnx
Description The verb noun pairs that select this tree family are pairs in which the interpre
tation is non comp ositional and the noun contributes argument structure to the predicate
e g Dania made Srini a cake vs Dania made Srini a loan The verb and the noun
CHAPTER VERB CLASSES
Sr
NP0↓ VP
V◊ NP1
N1◊
Figure Declarative LightVerb Tree nx lVN
o ccur together in the syntactic database and b oth anchor the trees The verbs in these
lightverb constructions are give and make The noun following the lightverb is usually
a bare in nitive form e g make a promise to Anoop However we include deverbal
nominals e g make a payment to Anoop aswell Constructions with nouns that do not
contribute an argument structure are excluded In addition to semantic considerations
of light verbs they di er syntactically from the Ditransitive with PP Shift verbs see
section as w ell in that the noun in the lightverb construction do es not extract Also
passivization is severely restricted Sp ecial determiner requirments and restrictions are
handled in the same manner as for the Tnx lVN family There are verb noun pairs
that select this family
Examples give look give wave make promise
Dania gave Carl a murderous look
Amanda gave us a little wave as she left
Dania made Doug a promise
Declarative tree See Figure
S r Sr
NP0↓ VP NP0↓ VP
V◊ NP1 PP2
V◊ NP2↓ NP1 N1◊ P2 NP2↓
to N1◊
a b
Figure Declarative LightVerbs with PP Tree nx lVN Pnx a nx lVnx N b
Other available trees Non shifted wh moved sub ject wh moved indirect ob ject sub ject
relative clause with and without comp adjunct gap less relative clause with comp with
PP pied piping indirect ob ject relative clause with and without comp with PP pied
piping imp erative NP gerund passive with by phrase passive with by phrase with ad
junct gap less relative clause with comp with PP pied piping gerund passive with by
phrase gerund passive without by phrase
Shifted wh moved sub ject wh moved ob ject of PP wh moved PP sub ject relative
clause with and without comp ob ject of PP relative clause with and without comp with
PP pied piping imp erative determiner gerund NP gerund passive with by phrase with
adjunct gap less relative clause with comp with PP pied piping gerund passive with by
phrase gerund passive without by phrase
NP It Cleft TItVnx s
Description This tree family is selected by be as the main verb and it as the sub ject Together
these two items serve as a multi comp onent anchor for the tree family This tree family
is used for it clefts in which the clefted element is an NP and there are no gaps in the
clause which follows the NP The NP is interpreted as an adjunct of the following clause
See Chapter for additional discussion
Examples it be
it was yesterday that we had the meeting
Declarative tree See Figure
Sr
NP0 VP
N◊ V◊ VP1
V1 PP1 S2↓
ε P1↓ NP1↓
Figure Declarative NP It Cleft Tree ItVpnx s
Other available trees inverted question wh moved ob ject with be inverted wh moved ob
ject with be not inverted adjunct gap less relative clause with comp with PP pied
piping
PP It Cleft TItVpnx s
Description This tree family is selected by be as the main verb and it as the sub ject Together
these two items serveasamulti comp onentanchor for the tree family This tree family is
CHAPTER VERB CLASSES
used for it clefts in which the clefted element is a PP and there are no gaps in the clause
which follows the PP The PP is interpreted as an adjunct of the following clause See
Chapter for additional discussion
Examples it be
it was at Kent State that the police shot al l those students
Declarative tree See Figure
Sr
NP0 VP
N◊ V◊ VP1
V1 PP1 S2↓
ε P1↓ NP1↓
Figure DeclarativePPIt CleftTree ItVnx s
Other available trees inverted question wh moved prep ositional phrase with be inverted
wh moved prep ositional phrase with be not inverted adjunct gap less relative clause
with comp with PP pied piping
Adverb It Cleft TItVad s
Description This tree family is selected by be as the main verb and it as the sub ject Together
these two items serve as a multi comp onent anchor for the tree family This tree family
is used for it clefts in which the clefted elementisanadverb and there are no gaps in the
clause which follows the adverb The adverbisinterpreted as an adjunct of the following
clause See Chapter for additional discussion
Examples it be
it was reluctantly that Dania agreed to do the tech report
Declarative tree See Figure
Other available trees inverted question wh moved adverb how with be inverted wh moved
adverb how with be not inverted adjunct gap less relative clause with comp with PP
pied piping
Adjective Small Clause Tree Tnx Ax
Description These trees are not anchored by verbs but by adjectives They are explained Sr
NP0 VP
N◊ V◊ VP1
V1 Ad1↓ S2↓
ε
Figure Declarative Adverb It Cleft Tree ItVad s
in much greater detail in the section on small clauses see section This section is
presented here for completeness adjectives select this tree family
Examples addictive dangerous wary
cigarettes are addictive
smoking cigarettes is dangerous
John seems wary of the Surgeon General s warnings
Declarative tree See Figure
Sr
NP0↓ VP
V AP1 NA
ε A◊
Figure Declarative Adjective Small Clause Tree nx Ax
Other available trees wh moved sub ject wh moved adjective how relative clause on sub ject
with and without comp imp erative NP gerund adjunct gap less relative clause with
comp with PP pied piping
Adjective Small Clause with Sentential Complement Tnx A s
Description This tree family is selected by adjectives that takesentential complements The
sentential complements can b e indicative or in nitive Note that these trees are anchored
by adjectives not verbs Small clauses are explained in much greater detail in section
This section is presented here for completeness adjectives select this tree family
CHAPTER VERB CLASSES
Examples able curious disappointed
Christy was able to nd the problem
Christy was curious whether the new analysis was working
Christy was sad that the old analysis failed
Declarative tree See Figure
Sr
NP0↓ VP
V AP1 NA
ε A◊ S1↓
Figure Declarative Adjective Small Clause with Sentential ComplementTree nx A s
Other available trees wh moved sub ject wh moved adjective how relative clause on sub ject
with and without comp imp erative NP gerund adjunct gap less relative clause with
comp with PP pied piping
Adjective Small Clause with Sentential Sub ject Ts Ax
Description This tree family is selected by adjectives that take sentential sub jects The
sentential sub jects can be indicativeor in nitive Note that these trees are anchored by
adjectives not verbs Most adjectives that take the Adjective Small Clause tree family
see section take this family as well Small clauses are explained in much greater
detail in section This section is presented here for completeness adjectives
select this tree family
Examples decadent incredible uncertain
to eat raspberry chocolate tru e icecream is decadent
that Carl could eat a large bow l of it is incredible
whether he wil l actual ly survive the experience is uncertain
Declarative tree See Figure
vailable trees wh moved sub ject wh moved adjective adjunct gap less relative Other a
clause with comp with PP pied piping
No great attempt has b een made to go through and decide which adjectives should actually take this family
and which should not Sr
S0↓ VP
V AP1 NA
ε A◊
Figure Declarative Adjective Small Clause with Sentential Sub ject Tree s Ax
Equative BE Tnx BEnx
Description This tree family is selected only by the verb be It is distinguished from the
predicative NP s see section in that two NP s are equated and hence interchange
able see Chapter for more discussion on the English copula and predicative sentences
The XTAG analysis for equative be is explained in greater detail in section
Examples be
That man is my uncle
Declarative tree See Figure
Sr
NP0↓ VPr
V◊ VP1
V1 NP1↓
ε1
Figure DeclarativeEquative BE Tree nx BEnx
Other available trees inverted question
NP Small Clause Tnx N
Description The trees in this tree family are not anchored by verbs but by nouns Small
clauses are explained in much greater detail in section This section is presented here
for completeness nouns select this tree family
CHAPTER VERB CLASSES
Examples author chair dish
Dania is an author
that blue warped looking thing is a chair
those broken pieces were dishes
Declarative tree See Figure
Sr
NP0↓ VP
V NP1
ε N◊
Figure Declarative NP Small Clause Trees nx N
Other available trees wh moved sub ject wh moved ob ject relative clause on sub ject with
and without comp imp erative NP gerund adjunct gap less relative clause with comp with
PP pied piping
NP Small Clause with Sentential Complement Tnx N s
Description This tree family is selected by the small group of nouns that takesentential com
plements by themselves see section The sentential complements can b e indicativeor
in nitive dep ending on the noun Small clauses in general are explained in much greater
detail in the section This section is presented here for completeness nouns select
this family
Examples admission claim vow
The a davits are admissions that they kil led the sheep
there is always the claim that they were insane
this is his vow to ght the charges
Declarative tree See Figure
Other available trees wh moved sub ject wh moved ob ject relative clause on sub ject with
and without comp imp erative NP gerund adjunct gap less relative clause with comp with
PP pied piping
NP Small Clause with Sentential Sub ject Ts N
Description This tree family is selected by nouns that takesentential sub jects The sentential
sub jects can b e indicative or in nitive Note that these trees are anchored by nouns not Sr
NP0↓ VP
V NP1 NA
ε N◊ S1↓
Figure DeclarativeNPwithSentential Complement Small Clause Tree nx N s
verbs Most nouns that take the NP Small Clause tree family see section take this
family as well Small clauses are explained in much greater detail in section This
section is presented here for completeness nouns select this tree family
Examples dilemma insanity tragedy
whether to keep the job he hates is a dilemma
to invest al l of your money in worms is insanity
that the worms died is a tragedy
Declarative tree See Figure
Sr
S0↓ VP
V NP1 NA
ε N◊
Figure Declarative NP Small Clause with Sentential Sub ject Tree s N
Other available trees wh moved sub ject adjunct gap less relative clause with comp with
PP pied piping
PP Small Clause Tnx Pnx
Description This family is selected by prep ositions that can o ccur in small clause construc
tions For more information on small clause constructions see section This section
is presented here for completeness prep ositions select this tree family
No great attempt has b een made to go through and decide which nouns should actually take this family and
which should not
CHAPTER VERB CLASSES
Examples around in underneath
Chris is around the corner
Trisha is in big trouble
The dog is underneath the table
Declarative tree See Figure
Sr
NP0↓ VP
V PP1 NA
ε P◊ NP1↓
Figure Declarative PP Small Clause Tree nx Pnx
Other available trees wh moved sub ject wh moved ob ject of PP relative clause on sub ject
with and without comp relative clause on ob ject of PP with and without comp with PP
pied piping imp erative NP gerund adjunct gap less relative clause with comp with
PP pied piping
Exhaustive PP Small Clause Tnx Px
Description This family is selected by exhaustive prep ositions that can occur in small
clauses Exhaustive prep ositions are prep ositions that function as prep ositional phrases
by themselves For more information on small clause constructions please see section
The section is included here for completeness exhaustive prep ositions select this tree
family
Examples abroad below outside
Dr Joshi is abroad
The workers are al l below
Clove is outside
Declarative tree See Figure
Other available trees wh moved sub ject wh moved PP relative clause on sub ject with and
without comp imp erative NP gerund adjunct gap less relative clause with comp with
PP pied piping
PP Small Clause with Sentential Sub ject Ts Pnx
Description This tree family is selected by prep ositions that take sentential sub jects The
sentential sub ject can be indicative or in nitive Small clauses are explained in much Sr
NP0↓ VP
V PP1 NA
ε P◊
Figure Declarative Exhaustive PP Small Clause Tree nx Px
greater detail in section This section is presented here for completeness prep osi
tions select this tree family
Examples beyond unlike
that Ken could forget to pay the taxes is beyond belief
to explain how this happened is outside the scope of this discussion
for Ken to do something right is unlike him
Declarative tree See Figure
Sr
S0↓ VP
V PP1 NA
ε P◊ NP1↓
Figure Declarative PP Small Clause with Sentential Sub ject Tree s Pnx
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp with PP pied piping adjunct gap less relative clause with comp with PP
pied piping
Intransitive Sentential Sub ject Ts V
Description Only the verb matter selects this tree family The sentential sub ject can be
indicative complementizer required or in nitive complementizer optional
Examples matter
to arrive on time matters considerably
that Joshi attends the meetings matters to everyone
Declarative tree See Figure
CHAPTER VERB CLASSES
Sr
S0↓ VP
V◊
Figure DeclarativeIntransitiveSentential Sub ject Tree s V
Other available trees wh moved sub ject adjunct gap less relative clause with comp with
PP pied piping
Sentential Sub ject with to complement Ts Vtonx
Description The verbs that select this tree family are fal l occur and leak The sentential sub
ject can b e indicative complementizer required or in nitive complementizer optional
Examples fal l occur leak
to wash the car fel l to the children
that he should leave occurred to the party crasher
whether the princess divorced the prince leaked to the press
Declarative tree See Figure
Sr
S0↓ VP
V◊ PP1
P1 NP1↓
to
Figure Sentential Sub ject Tree with to complement s Vtonx
Other available trees wh moved sub ject adjunct gap less relative clause with comp with
PP pied piping
PP Small Clause with Adv and Prep anchors Tnx ARBPnx
Description This family is selected bymulti word prep ositions that can o ccur in small clause
constructions In particular this family is selected by two word prep ositions where the
rst word is an adverb the second word a prep osition Both comp onents of the multi
word prep osition are anchors For more information on small clause constructions see
section multi word prep ositions select this tree family
Examples ahead of close to
The little girl is ahead of everyone else in the race
The project is close to completion
Declarative tree See Figure
Sr
NP0↓ VP
V PP1
ε P1 NP1↓
Ad◊ P◊
Figure Declarative PP Small Clause tree with two word prep osition where the rst word
is an adverb and the second word is a prep osition nx ARBPnx
Other available trees wh moved sub ject wh moved ob ject of PP relative clause on sub ject
with and without comp relative clause on ob ject of PP with and without comp adjunct
gap less relative clause with comp with PP pied piping imp erative NP Gerund
PP Small Clause with Adj and Prep anchors Tnx APnx
Description This family is selected bymulti word prep ositions that can o ccur in small clause
constructions In particular this family is selected by two word prep ositions where the
rst word is an adjective the second word a prep osition Both comp onents of the multi
word prep osition are anchors For more information on small clause constructions see
section multi word prep ositions select this tree family
Examples according to void of
standard procedure The operation we performed was according to
He is void of al l feeling
Declarative tree See Figure
Other available trees wh moved sub ject relative clause on sub ject with and without comp
relative clause on ob ject of PP with and without comp wh moved ob ject of PP adjunct
gap less relative clause with comp with PP pied piping
CHAPTER VERB CLASSES
Sr
NP0↓ VP
V PP1 NA
ε P1 NP1↓
A◊ P◊
Figure Declarative PP Small Clause tree with two word prep osition where the rst word
is an adjective and the second word is a prep osition nx APnx
PP Small Clause with Noun and Prep anchors Tnx NPnx
Description This family is selected bymulti word prep ositions that can o ccur in small clause
constructions In particular this family is selected by two word prep ositions where the
rst word is a noun the second word a prep osition Both comp onents of the multi
word prep osition are anchors For more information on small clause constructions see
section multi word prep osition selects this tree family
Examples thanks to
The fact that we are here tonight is thanks to the valiant e orts of our sta
Declarative tree See Figure
Sr
NP0↓ VP
V PP1 NA
ε P1 NP1↓
N◊ P◊
Figure Declarative PP Small Clause tree with two word prep osition where the rst word
is a noun and the second word is a prep osition nx NPnx
Other available trees wh moved sub ject wh moved ob ject of PP relative clause on sub ject
with and without comp relative clause on ob ject with comp adjunct gap less relative
clause with comp with PP pied piping
PP Small Clause with Prep anchors Tnx PPnx
Description This family is selected bymulti word prep ositions that can o ccur in small clause
constructions In particular this family is selected bytwo word prep ositions where b oth
words are prep ositions Both comp onents of the multi word prep osition are anchors For
more information on small clause constructions see section multi word prep ositions
select this tree family
Examples on to inside of
that detective is on to you
The red box is inside of the blue box
Declarative tree See Figure
Sr
NP0↓ VP
V PP1 NA
ε P NP1↓
P1◊ P2◊
Figure Declarative PP Small Clause tree with two word prep osition where b oth words
are prep ositions nx PPnx
Other available trees wh moved sub ject wh moved ob ject of PP relative clause on sub ject
with and without comp relative clause on ob ject of PP with and without comp with PP
pied piping imp erative wh moved ob ject of PP adjunct gap less relative clause with
comp with PP pied piping
PP Small Clause with Prep and Noun anchors Tnx PNaPnx
Description This family is selected bymulti word prep ositions that can o ccur in small clause
constructions In particular this family is selected by three word prep ositions The rst
and third words are always prep ositions and the middle word is a noun The noun is
y noun mo di ers All three marked for null adjunction since it cannot be mo di ed b
comp onents of the multi word prep osition are anchors For more information on small
clause constructions see section multi word prep osition select this tree family
CHAPTER VERB CLASSES
Examples in back of in line with on top of
The red plaid box should be in back of the plain black box
The evidence is in line with my new ly concocted theory
She is on top of the world
She is on direct top of the world
Declarative tree See Figure
Sr
NP0↓ VP
V PP1 NA
ε P NP1↓
P1◊ N◊ P2◊
NA
Figure Declarative PP Small Clause tree with three word prep osition where the middle
noun is marked for null adjunction nx PNaPnx
Other available trees wh moved sub ject wh moved ob ject of PP relative clause on sub ject
with and without comp relative clause on ob ject of PP with and without comp with PP
pied piping adjunct gap less relative clause with comp with PP pied piping imp erative
NP Gerund
PP Small Clause with Sentential Sub ject and Adv and
Prep anchors Ts ARBPnx
Description This tree family is selected bymulti word prep ositions that takesentential sub
jects In particular this family is selected bytwo word prep ositions where the rst word
is an adverb the second word a prep osition Both comp onents of the multi word prep osi
tion are anchors The sentential sub ject can b e indicative or in nitive Small clauses are
explained in much greater detail in section prep ositions select this tree family
Examples due to contrary to
never sleeps during the week that David slept until noon is due to the fact that he
that Michael s joke was funny is contrary to the usual status of his comic attempts
Declarative tree See Figure Sr
S0↓ VP
V PP1 NA
ε P1 NP1↓
Ad◊ P◊
Figure Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where the rst wordisanadverb and the second word is a prep osition s ARBPnx
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp adjunct gap less relative clause with comp with PP pied piping
PP Small Clause with Sentential Sub ject and Adj and
Prep anchors Ts APnx
Description This tree family is selected bymulti word prep ositions that takesentential sub
jects In particular this family is selected bytwo word prep ositions where the rst word
is an adjective the second word a prep osition Both comp onents of the multi word prep o
sition are anchors The sentential sub ject can be indicative or in nitive Small clauses
are explained in much greater detail in section prep ositions select this tree family
Examples devoid of according to
that he could walk out on her is devoid of al l reason
that the conversation erupted precisely at that moment was according to my theory
Declarative tree See Figure
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp adjunct gap less relative clause with comp with PP pied piping
PP Small Clause with Sentential Sub ject and Noun and
Prep anchors Ts NPnx
Description This tree family is selected bymulti word prep ositions that takesentential sub
jects In particular this family is selected bytwo word prep ositions where the rst word
is a noun the second word a prep osition Both comp onents of the multi word prep osi
CHAPTER VERB CLASSES
Sr
S0↓ VP
V PP1 NA
ε P1 NP1↓
A◊ P◊
Figure Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where the rst word is an adjective and the second word is a prep osition s APnx
tion are anchors The sentential sub ject can b e indicative or in nitive Small clauses are
explained in much greater detail in section prep osition selects this tree family
Examples thanks to
that she is worn out is thanks to a long day in front of the computer terminal
Declarative tree See Figure
Sr
S0↓ VP
V PP1 NA
ε P1 NP1↓
N◊ P◊
Figure Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where the rst wordisanoun andthesecondword is a prep osition s NPnx
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp adjunct gap less relative clause with comp with PP pied piping
PP Small Clause with Sentential Sub ject and Prep an
chors Ts PPnx
Description This tree family is selected bymulti word prep ositions that takesentential sub
jects In particular this family is selected bytwo word prep ositions where b oth words are
prep ositions Both comp onents of the multi word prep osition are anchors The sentential
sub ject can b e indicative or in nitive Small clauses are explained in much greater detail
in section prep ositions select this tree family
Examples outside of
that Mary did not complete the task on time is outside of the scope of this discussion
Declarative tree See Figure
Sr
S0↓ VP
V PP1 NA
ε P NP1↓
P1◊ P2◊
Figure Declarative PP Small Clause with Sentential Sub ject Tree with two word prep o
sition where b oth words are prep ositions s PPnx
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp adjunct gap less relative clause with comp with PP pied piping
PP Small Clause with Sentential Sub ject and Prep and
Noun anchors Ts PNaPnx
Description This tree family is selected bymulti word prep ositions that take sentential sub
jects In particular this family is selected by three word prep ositions The rst and third
words are always prep ositions and the middle word is a noun The noun is marked for
null adjunction since it cannot be mo di ed by noun mo di ers All three comp onents
Small clauses are explained in much greater of the multi word prep osition are anchors
detail in section prep ositions select this tree family
Examples on account of in support of
that Joe had to leave the beach was on account of the hurricane
CHAPTER VERB CLASSES
that Maria could not come is in support of my theory about her
that Maria could not come is in direct strict desparate support of my theory about her
Declarative tree See Figure
Sr
S0↓ VP
V PP1 NA
ε P NP1↓
P1◊ N◊ P2◊
NA
Figure Declarative PP Small Clause with Sentential Sub ject Tree with three word prep o
sition where the middle noun is marked for null adjunction s PNaPnx
Other available trees wh moved sub ject relative clause on ob ject of the PP with and with
out comp adjunct gap less relative clause with comp with PP pied piping
Predicative Adjective with Sentential Sub ject and Com
plement Ts A s
Description This tree family is selected by predicative adjectives that take sentential sub jects
and a sentential complement This tree family is selected by likely and certain
Examples likely certain
that Max continues to drive a Jaguar is certain to make Bil l jealous
for the Jaguar to be towed seems likely to make Max very angry
Declarative tree See Figure
ailable trees wh moved sub ject adjunct gap less relative clause with comp with Other av
PP pied piping
Lo cative Small Clause with Ad anchor Tnx nx ARB
Description These trees are not anchored byverbs but byadverbs that are part of lo cative
adverbial phrases Lo catives are explained in much greater detail in the section on the
lo cative mo di er trees see section The only remarkable asp ect of this tree family is Sr
S0↓ VP
V AP1 NA
ε A◊ S1↓
Figure Predicative AdjectivewithSentential Sub ject and Complement s A s
the wh moved lo cativetree W nx nx ARB shown in Figure This is the only tree
family with this typ e of transformation in which the entire adverbial phrase is wh moved
but not all elements are replaced by wh items as in how many city blocks away is the
record store Lo catives that consist of just the lo cative adverb or the lo cative adverb
and a degree adverb see Section for details are treated as exhaustive PPs and
therefore select that tree family Section when used predicatively For an extensive
description of small clauses see Section adverbs select this tree family
Examples ahead o shore behind
the crash is three blocks ahead
the naval battle was many kilometers o shore
how many blocks behind was Max
Declarative tree See Figure
Sr
NP0↓ VP
V AdvP NA
ε NP1↓ Ad◊
Figure DeclarativeLocative Adverbial Small Clause Tree nx nx ARB
Other available trees wh moved sub ject relative clause on sub ject with and without comp
wh moved lo cative imp erative NP gerund
CHAPTER VERB CLASSES
S q
AdvP S r
NP ↓ Ad◊ NP 0↓ VP
V AdvP 1 NA
ε0 ε
Figure Wh moved Lo cative Small Clause Tree W nx nx ARB
Exceptional Case Marking TXnx Vs
Description This tree family is selected byverbs that are classi ed as exceptional case mark
ing meaning that the verb asssigns accusative case to the sub ject of the sentential com
plement This is in contrast to verbs in the Tnx Vnx s family section which assign
accusative case to a NP which is not part of the sentential complement ECM verbs take
sentential complements whichare either an in nitiveor a bare in nitive As with the
Tnx Vs family section the declarative and other trees in the Xnx Vs family are
auxiliary trees as opp osed to the more common initial trees These auxiliary trees adjoin
onto an S no de in an existing tree of the typ e sp eci ed by the sentential complement
This is the mechanism by which TAGs are able to maintain long distance dep endencies
see Chapter even over multiple emb eddings e g Who did Bil l expect to eat beans
or who did Bil l expect Mary to like See section for details on this family verbs
select this tree family
Examples expect see
Van expects Bob to talk Bob sees the harmonica fal l
Declarative tree See Figure
Sr
NP0↓ VP
V◊ S1*
Figure ECM Tree Xnx Vs
Other available trees wh moved sub ject sub ject relative clause with and without comp
adjunct gap less relative clause with and without comp with PP pied piping imp erative
NP gerund
Idiom with V D and N anchors Tnx VDN
Description This tree family is selected by idiomatic phrases in which the verb determiner
and NP are all frozen as in He kicked the bucket Only a limited number of transfor
mations are allowed as compared to the normal transitive tree family see section
Other idioms that have the same structure as kick the bucket and that are limited to
the same transformations would select this tree while di erent tree families are used to
handle other idioms Note that John kicked the bucket is actually ambiguous and would
result in two parses an idiomatic one meaning that John died and a comp ositional
transitive one meaning that there is an physical bucket that John hit with his fo ot
idiom selects this family
Examples kick the bucket
Nixon kicked the bucket
Declarative tree See Figure
Sr
NP0↓ VP
V◊ NP1
DetP1 N1◊
D1◊
Figure Declarative Transitive Idiom Tree nx VDN
Other available trees sub ject relative clause with and without comp declarative wh moved
sub ject imp erative NP gerund adjunct gapless relative with comp with PP pied piping
passive w wo by phrase wh moved ob ject of by phrase wh moved by phrase relative
with and without comp on sub ject of passive PP relative
Idiom with V D A and N anchors Tnx VDAN
Description This tree family is selected by transitive idioms that are anchored by a verb
determiner adjective and noun idioms select this family
CHAPTER VERB CLASSES
Examples have a green thumb sing a di erent tune
Martha might have a green thumb but it s uncertain after the death of al l the plants
After his conversion John sang a di erent tune
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ NP1
D1 ◊ A◊ N1 ◊
Figure Declarative Idiom with V D A and N Anchors Tree nx VDAN
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase passive with relativeonobjectofby phrase with and without
comp
Idiom with V and N anchors Tnx VN
Description This tree family is selected by transitive idioms that are anchored bya verb and
noun idioms select this family
Examples draw blood cry wolf
Graham s retort drew blood
The neglected boy cried wolf
Declarative tree See Figure
trees Sub ject relative clause with and without comp adjunct relative clause Other available
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase passive with relativeonobjectofby phrase with and without
comp Sr
NP0 ↓ VP
V◊ NP1
N1 ◊
Figure Declarative Idiom with V and N Anchors Tree nx VN
Idiom with V A and N anchors Tnx VAN
Description This tree family is selected by transitive idioms that are anchored by a verb
adjective and noun idioms select this family
Examples break new ground cry bloody murder
The avant garde lm breaks new ground
The investors cried bloody murder after the suspicious takeover
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ NP1
A◊ N1 ◊
Figure Declarative Idiom with V A and N Anchors Tree nx VAN
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase passive with relativeonobjectofby phrase with and without
comp
CHAPTER VERB CLASSES
Idiom with V D A N and Prep anchors Tnx VDAN Pnx
Description This tree family is selected by transitive idioms that are anchored by a verb
determiner adjective noun and prep osition idioms select this family
Examples make a big deal about make a great show of
John made a big deal about a miniscule dent in his car
The company made a big show of paying generous dividends
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ NP1 PP2
D1 ◊ A◊ N1 ◊ P2 ◊ NP2 ↓
Figure Declarative Idiom with V D A N and Prep Anchors Tree nx VDAN Pnx
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase outer passive without by phrase outer passivewithby phrase
outer passive with wh moved by phrase outer passive with wh moved ob ject of by phrase
outer passive without by phrase with relative on the sub ject with and without comp outer
passivewithby phrase with relative on sub ject with and without comp
Idiom with V A N and Prep anchors Tnx VAN Pnx
Description This tree family is selected by transitive idioms that are anchored by a verb
adjective noun and prep osition idioms select this family
Examples make short work of
John made short work of the glazed ham
Declarative tree See Figure
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive Sr
NP0 ↓ VP
V◊ NP1 PP2
A◊ N1 ◊ P2 ◊ NP2 ↓
Figure Declarative Idiom with V A N and Prep Anchors Tree nx VAN Pnx
with wh moved by phrase outer passive without by phrase outer passivewithby phrase
outer passive with wh moved by phrase outer passive with wh moved ob ject of by phrase
outer passive without by phrase with relative on the sub ject with and without comp outer
passivewithby phrase with relative on sub ject with and without comp
Idiom with V N and Prep anchors Tnx VN Pnx
Description This tree family is selected by transitive idioms that are anchored by a verb
noun and prep osition idioms select this family
Examples look daggers at keep track of
Maria looked daggers at her ex husband across the courtroom
The company kept track of its inventory
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ NP1 PP2
N1 ◊ P2 ◊ NP2 ↓
Figure Declarative Idiom with V N and Prep Anchors Tree nx VN Pnx
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
CHAPTER VERB CLASSES
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase outer passive without by phrase outer passivewithby phrase
outer passive with wh moved by phrase outer passive with wh moved ob ject of by phrase
outer passive without by phrase with relative on the sub ject with and without comp outer
passivewithby phrase with relative on sub ject with and without comp
Idiom with V D N and Prep anchors Tnx VDN Pnx
Description This tree family is selected by transitive idioms that are anchored by a verb
determiner noun and prep osition idioms select this family
Examples make a mess of keep the lid on
John made a mess out of his new suit
The tabloid didn t keep a lid on the imminent celebrity nuptials
Declarative tree See Figure
Sr
NP0 ↓ VP
V◊ NP1 PP2
D1 ◊ N1 ◊ P2 ◊ NP2 ↓
Figure Declarative Idiom with V D N and Prep Anchors Tree nx VDN Pnx
Other available trees Sub ject relative clause with and without comp adjunct relative clause
with comp with PP pied piping wh moved sub ject imp erative NP gerund passive with
out by phrase passivewithby phrase passive with wh moved ob ject of by phrase passive
with wh moved by phrase outer passive without by phrase outer passivewithby phrase
outer passive with wh moved by phrase outer passive with wh moved ob ject of by phrase
outer passive without by phrase with relative on the sub ject with and without comp outer
passivewithby phrase with relative on sub ject with and without comp
Chapter
Ergatives
Verbs in English that are termed ergative display the kind of alternation shown in the sentences
in b elow
The sun melted the ice
The ice melted
The pattern of ergative pairs as seen in is for the ob ject of the transitive sentence to
b e the sub ject of the intransitivesentence The literature discussing such pairs is based largely
on syntactic mo dels that involvemovement particularly GB Within that framework two basic
approaches are discussed
Derived Intransitive
The intransitive member of the ergative pair is derived through pro cesses of movement
and deletion from
a transitive D structure Burzio or
transitive lexical structure Hale and Keyser Hale and Keyser
Pure Intransitive
The intransitivememberisintransitiv eat all levels of the syntax and the lexicon and is
not related to the transitivememb er syntactically or lexically Nap oli
The Derived Intransitive approach s notions of movement in the lexicon or in the grammar
are not represented as such in the XTAG grammar However distinctions drawn in these ar
guments can b e translated to the FB LTAG framework In the XTAG grammar the di erence
between these two approaches is not a matter of movement but rather a question of tree fam
ily membership The relation between sentences represented in terms of movement in other
frameworks is represented in XTAGbymemb ership in the same tree family Wh questions and
their indicativecounterparts are one example of this Adopting the Pure Intransitive approach
suggested by Nap oli would mean placing the intransitive ergatives in a tree family with
other intransitiveverbs and separate from the transitivevariants of the same verbs This would
result in a grammar that represented intransitive ergatives as more closely related to other
intransitives than to their transitive counterparts The only hint of the relation between the
CHAPTER ERGATIVES
intransitive ergatives and the transitive ergatives would be that ergative verbs would select
b oth tree families While this is a workable solution it is an unattractive one for the English
XTAG grammar b ecause semantic coherence is implicitly asso ciated with tree families in our
analysis of other constructions In particular constancy in thematic role is represented by
constancy in no de names across sentence typ es within a tree family For example if the ob ject
of a declarative tree is NP the sub ject of the passive tree s in that family will also b e NP
The analysis that has been implemented in the English XTAG grammar is an adaptation
of the Derived Intransitive approach The ergative verbs select one family Tnx Vnx that
contains b oth transitive and intransitive trees The trans feature app ears on the intransitive
ergative trees with the value and on the transitive trees with the value This creates the
two p ossibilities needed to account for the data
intransitive ergative transitive alternation These verbs have transitive and intran
sitivevariants as shown in sentences and
The sun melted the ice cream
The ice cream melted
In the English XTAG grammar verbs with this b ehavior are left unsp eci ed as to value
for the trans feature This lack of sp eci cation allows these verbs to anchor either
typ e of tree in the Tnx Vnx tree family b ecause the unsp eci ed trans value of the
verb can unify with either or values in the trees
transitive only Verbs of this typ e select only the transitive trees and do not allow
intransitiveergative variants as in the pattern show in sentences and
Elmo b orrowed a book
A b o ok b orrowed
The restriction to selecting only transitive trees is accomplished by setting the trans
feature value to for these verbs
Sr
NP1↓ VP
V◊ trans : -
Figure Ergative Tree Enx V
The declarative ergative tree is shown in Figure with the trans feature displayed
Note that the index of the sub ject NP indicates that it originated as the ob ject of the verb
Chapter
Sentential Sub jects and Sentential
Complements
In the XTAG grammar arguments of a lexical item including sub jects app ear in the initial tree
anchored by that lexical item Asentential argument app ears as an S no de in the appropriate
p osition within an elementary tree anchored by the lexical item that selects it This is the case
for sentential complements of verbs prep ositions and nouns and for sentential sub jects The
distribution of complementizers in English is intertwined with the distribution of emb edded
sentences A successful analysis of complementizers in English must handle b oth the co o ccur
rence restrictions between complementizers and various typ es of clauses and the distribution
of the clauses themselves in b oth sub ject and complement p ositions
S or VP complements
Two comparable grammatical formalisms Generalized Phrase Structure Grammar GPSG
Gazdar et al and Head driven Phrase Structure Grammar HPSG Pollard and Sag
have rather di erent treatments of sentential complements S comps They b oth treat
emb edded sentences as VP s with sub jects which generates the correct structures but misses
the generalization that S s behave similarly in both matrix and emb edded environments and
VP s b ehave quite di erently Neither account has PRO sub jects of in nitival clauses they
have sub jectless VP s instead GPSG has a complete complementizer system which app ears
to cover the same range of data as our analysis It is not clear what sort of complementizer
analysis could b e implemented in HPSG
Following standard GB approach the English XTAG grammar do es not allow VP com
plements but treats verb anchored structures without overt sub jects as having PRO sub jects
Thus indicative clauses in nitives and gerunds all have a uniform treatment as emb edded
clauses using the same trees under this approach Furthermore our analysis is able to preserve
the selectional and distributional distinction b etween S s and VP s in the spirit of GB theories
without having to posit extra empty categories Consider the alternation between that and
the null complementizer shown in sentences and
i e empty complementizers WedohavePRO and NP traces in the grammar
Although we will continue to refer to null complementizers in our analysis this is actually the absence of
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
He hop es Muriel wins
He hop es that Muriel wins
In GB both Muriel wins in and that Muriel wins in are CPs even though there
is no overt complementizer to head the phrase in Our grammar do es not distinguish by
category lab el b etween the phrases that would b e lab eled in GB as IP and CP We lab el b oth
of these phrases S The di erence between these two levels is the presence or absence of the
complementizer or extracted WH constituent and is represented in our system as a di erence
in feature values here of the comp feature and the presence of the additional structure
contributed by the complementizer or extracted constituent This illustrates an imp ortan t
distinction in XTAG that b etween features and no de lab els Because wehave a sophisticated
feature system we are able to make ne grained distinctions b etween no des with the same lab el
which in another system mighthave to b e realized by using distinguishing no de lab els
Complementizers and Emb edded Clauses in English The
Data
Verbs selecting sentential complements or sub jects place restrictions on their complements
in particular on the form of the embedded verb phrase Furthermore complementizers are
constrained to app ear with certain typ es of clauses again based primarily on the form of the
emb edded VP For example hope selects both indicative and in nitival complements With
an indicative complement it may only have that or null as p ossible complementizers with an
in nitival complement it may only haveanull complementizer Verbs that allow wh comple
mentizers suchasask can take whether and if as complementizers The p ossible combinations
of complementizers and clause typ es is summarized in Table
As can be seen in Table sentential sub jects di er from sentential complements in re
quiring the complementizer that for all indicative and sub junctive clauses In sentential com
plements that often varies freely with a null complementizer as illustrated in
Christy hop es that Mike wins
Christy hop es Mike wins
Dania thinks that Newt is a liar
Dania thinks Newt is a liar
That Helms won so easily annoyed me
Helms won so easily annoyed me
a complementizer
Other considerations such as the relationship between the tense asp ect of the matrix clause and the
tense asp ect of a complement clause are also imp ortant but are not currently addressed in the currentEnglish
XTAG grammar
Complementizer that whether if for null
Clause typ e
indicative sub ject Yes Yes No No No
complement Yes Yes Yes No Yes
in nitive sub ject No Yes No Yes Yes
complement No Yes No Yes Yes
sub junctive sub ject Yes No No No No
complement Yes No No No Yes
gerundive complement No No No No Yes
base complement No No No No Yes
small clause complement No No No No Yes
Table Summary of Complementizer and Clause Combinations
Another fact which must be accounted for in the analysis is that in in nitival clauses
the complementizer for must app ear with an overt sub ject NP whereas a complementizer less
in nitival clause never has an overt sub ject as shown in See section for more
discussion of the case assignment issues relating to this construction
To lose would b e awful
For Penn to lose would b e awful
For to lose would b e awful
Penn to lose would b e awful
In addition some verbs select wh complements either questions or clauses with
whether or if Grimshaw
Jesse wondered who left
Jesse wondered if Barry left
Jesse wondered whether to leave
Jesse wondered whether Barry left
Jesse thought who left
Jesse thought if Barry left
Jesse thought whether to leave
Jesse thought whether Barry left
Most gerundive phrases are treated as NP s In fact all gerundive sub jects are treated as NP s and the only
gerundive complements which receiveasentential parse are those for which there is no corresp onding NP parse
This was done to reduce duplication of parses See Chapter for further discussion of gerunds
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
Features Required
As wehave seen ab ove clauses maybe wh or wh mayhave one of several com
plementizers or no complementizer and can be of various clause typ es The XTAG analysis
uses three features to capture these p ossibilities comp for the variation in complemen
tizers wh for the question vs non question alternation and mo de for clause typ es
In addition to these three features the assign comp feature represents complementizer
requirements of the emb edded v erb More detailed discussion of the assign comp feature
app ears b elow in the discussions of sentential sub jects and of in nitives The four features and
their p ossible values are shown in Table
Feature Values
comp that if whether for rel nil
mo de ind inf sub jnt ger base ppart nom prep
assign comp that if whether for rel ind nil inf nil
wh
Table Summary of RelevantFeatures
Distribution of Complementizers
Like other non arguments complementizers anchor an auxiliary tree shown in Figure
and adjoin to elementary clausal trees The auxiliary tree for complementizers is the only
alternative to having a complementizer p osition built into every sentential tree The latter
choice would mean having an empty complementizer substitute into every matrix sentence and
a complementizerless emb edded sentence to ll the substitution no de Our choice follows the
XTAG principle that initial trees consist only of the arguments of the anchor the S tree do es
for a complementizer and the COMP tree has only one argument an S not contain a slot
with particular features determined by the complementizer Complementizers select the typ e
of clause to which they adjoin through constraints on the mo de feature of the S fo ot no de
in the tree shown in Figure These features also pass up to the ro ot no de so that they are
visible to the tree where the emb edded sentence adjoins substitutes
The grammar handles the following complementizers that whether if for and no comple
mentizer and the clause typ es indicative in nitival gerundive past participial sub junctive
and small clause nom prep The comp feature in a clausal tree re ects the value of the
complementizer if one has adjoined to the clause
The comp and wh features receive their ro ot no de values from the particular com
plementizer whichanchors the tree The COMPs tree adjoins to an S no de with the feature
comp nil this feature indicates that the tree do es not already have a complementizer
adjoined to it We ensure that there are no stacked complementizers by requiring the fo ot
no de of COMPs to have comp nil
mo de actually con ates several typ es of information in particular verb form and mo o d
See section for a discussion of the di erence b etween complements and adjuncts in the XTAG grammar
Because ro ot S s cannot have complementizers the parser checks that the ro ot S has comp nil at the
end of the derivation when the S is also checked for a tensed verb Sc comp : <1> inv : - displ-const : <2> wh : <3> mode : <4> ind/sbjnct
Comp wh : <3> Sr* inv : - comp : <1> NA assign-comp : <1> wh : - wh : - comp : that mode : <4> comp : nil sub-conj : nil displ-const : <2> assign-comp : that
that
Figure Tree COMPs anchored by that
Case assignment for and the two to s
The assign comp feature is used to represent the requirements of particular typ es of clauses
for particular complementizers So while the comp feature represents constraints originat
ing from the VP dominating the clause the assign comp feature represents constraints
originating from the highest VP in the clause assign comp is used to control the the
app earance of sub jects in in nitival clauses see discussion of ECM constructions in to
blo ck bare indicative sentential sub jects bare in nitival sub jects are allowed and to blo ck
that trace violations
Examples and show that an accusative case sub ject is obligatory in an
in nitive clause if the complementizer for is present The in nitive clauses in is analyzed
in the English XTAG grammar as having a PRO sub ject
Christywants to pass the exam
Mikewants for her to pass the exam
Mikewants for she to pass the exam
Christywants for to pass the exam
The for to construction is particularly illustrative of the di culties and b ene ts faced in
using a lexicalized grammar It is commonly accepted that for b ehaves as a case assigning
complementizer in this construction assigning accusative case to the sub ject of the clause
erb do es not assign case to its sub ject p osition However in our featurized since the in nitival v
grammar the absence of a feature licenses anything so we must have overt null case assigned
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
by in nitives to ensure the correct distribution of PRO sub jects See section for more dis
cussion of case assignment This null case assignment clashes with accusative case assignment
if we simply add for as a standard complementizer since NP s including PRO are drawn from
the lexicon already marked for case Thus we must use the assign comp feature to pass
information ab out the verb up to the ro ot of the emb edded sentence To capture these facts
two in nitive to s are p osited One in nitive to has assign case none which forces a PRO
sub ject and assign comp inf nil which prevents for from adjoining The other in nitive
to has no value at all for assign case and has assign comp for ecm so that it can
only o ccur either with the complementizer for or with ECM constructions In those instances
either for or the ECM verb supplies the assign case value assigning accusative case to the
overt sub ject
Sentential Complements of Verbs
Tree families Tnx Vs Tnx Vnx s TItVnx s TItVpnx s TItVad s
Verbs that select sentential complements restrict the mo de and comp values for
those complements Since with very few exceptions long distance extraction is p ossible from
sentential complements the S complement no des are adjunction no des Figure shows the
declarative tree for sentential complements anchored by think
Sr
NP0↓ VP
V S1*
think
Figure Sentential complement tree nx Vs
The need for an adjunction no de rather than a substitution no de at S maynotbeobvious
until one considers the derivation of sentences with long distance extractions For example the
declarative in is derived by adjoining the tree in Figure b to the S no de of the tree in
Figure a Since there are no b ottom features on S the same nal result could have b een
achieved with a substitution no de at S
The emu thinks that the aardvark smells terrible
However adjunction is crucial in deriving sentences with long distance extraction as in
sentences and
Who do es the emu think smells terrible
For example long distance extraction is not p ossible from the S complement in it clefts Sr Sr
NP VP NP VP
DetP N V A DetP N V S1*
D aardvark smells terrible D emu thinks
the the
a b
Figure Trees for The emu thinks that the aardvark smel ls terrible
Who did the elephant think the panda heard the emusay smells terrible
The example in is derived from the trees for who smel ls terrible shown in Figure
and the emu thinks S shown in Figure b by adjoining the latter at the S no de of the
r
former This pro cess is recursive allowing sentences like Such a representation has b een
shown by Kro ch and Joshi to b e well suited for describing unb ounded dep endencies
Sq
NP Sr
N NP0 VP NA
who ε V A
smells terrible
Figure Tree for Who smel ls terrible
In English a complementizer may not app ear on a complement with an extracted sub ject
the that trace con guration This phenomenon is illustrated in
Which animal did the gira e saythathelikes
Which animal did the gira e say that likes him
Which animal did the gira e saylikes him
See Chapter for a discussion of do supp ort
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
These sentences are derived in XTAG by adjoining the tree for did the gira e say S at
the S no de of the tree for either which animal likes him to yield sentence or which
r
animal he likes to yield sentence That trace violations are blo cked by the presence
of the feature assign comp inf nil ind nil ecm feature on the bottom of the S no de
r
of trees with extracted sub jects W i e those used in sentences such as and
If a complementizer tree COMPs adjoins to a sub ject extraction tree at S its assign
r
comp that whether for if feature will clash and the derivation will fail If there is no
complementizer there is no feature clash and this will p ermit the derivation of sentences like
or of ECM constructions in which case the ECM verb will have assign comp ecm
see section for more discussion of the ECM case Complementizers may adjoin normally
to ob ject extraction trees such as those used in sentence and so ob ject extraction trees
havenovalue for the assign comp feature
In the case of indirect questions sub jacency follows from the principle that a given tree
cannot contain more than one wh element Extraction out of an indirect question is ruled out
b ecause a sentence like
Who do you wonder who e loves e
i j j i
would have to be derived from the adjunction of do you wonder into who who e loves e
i j j i
which is an ill formed elementary tree
Exceptional Case Marking Verbs
Tree family TXnx Vs Exceptional Case Marking verbs are those which assign accusative
tial complement This is in contrast to verbs in the Tnx Vnx s case to the sub ject of the senten
family section which assign accusative case to an NP which is not part of the sentential
complement
The sub ject of an ECM in nitive complement is assigned accusative case is a manner anal
ogoustothatofasubjectinafor to construction as describ ed in section As in the for to
case the ECM verb assigns accusativecaseinto the sub ject of the lower in nitive and so the in
nitive uses the to which has no value for assign case and has assign comp for ecm
The ECM verb has assign comp ecm and assign case acc on its fo ot The former
allows the assign comp features of the ECM verb and the to tree to unify and so b e used
together and the latter assigns the accusative case to the lower sub ject
Figure shows the declarative tree for the tree for the TXnx Vs family in this case
Figure shows a parse for Van expects Bob to talk anchored by expects
The ECM and for to cases are analogous in how they are used together with the correct
in nitival to to assign accusative case to the sub ject of the lower in nitive However they are
di erent in that for is blo cked along with other complementizers in sub ject extraction contexts
as discussed in section as in while sub ject extraction is compatible with ECM cases
as in
What child did the gira e ask for to leave
This do es not mean that elementary trees with more than one gap should b e ruled out across the grammar
Such trees might b e required for dealing with parasitic gaps or gaps in co ordinated structures Sr
NP0 ↓ VP
V S1 * displ-const : set1 : - assign-comp : ecm inv : - extracted : - control : <2> punct : contains : <10> comp: nil mode: inf wh : -
expects
Figure ECM tree Xnx Vs
Sr
NP VP
N V S1
Van expects NP VPr
N V VP NA
Bob to V
talk
Figure Sample ECM parse
Who did Bill exp ect to eat b eans
nil ind nil ecm feature on the Sentence is ruled out bythe assign comp inf
sub ject extraction tree for ask since the assign comp for feature from the for tree will
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
fail to unify However will b e allowed since assign comp ecm feature on the expect
tree will unify with the fo ot of the ECM verb tree The use of features allows the ECM and for
to constructions to act the same for exceptional case assignment while also b eing distinguished
for that trace violations
Verbs that take bare in nitives as in are also treated as ECM verbs the only di erence
b eing that their fo ot feature has mo de base instead of mo de inf Since the comple
mentdoesnothave to there is no question of using the to tree for allowing accusative case to b e
assigned Instead verbs with mo de base allow either accusative or nominative case to b e
assigned to the sub ject and the fo ot of the ECM bare in nitive tree forces it to b e accusative
byits assign case acc value at its fo ot no de uni es with the assign case nom acc
value of the bare in nitive clause
Bob sees the harmonica fall
The trees in the TXnx Vs family are generally parallel to those in the Tnx Vs family
except for the assign case and assign comp values on the fo ot no des However the
TXnx Vs family also includes a tree for the passive which of course is not included in the
Tnx Vs family Unlike all the other trees in the TXnx Vs family the passive tree is not ro oted
in S and is instead a VP auxiliary tree Since the sub ject of the in nitiveisnot thematically
selected by the ECM verb it is not part of the ECM verb s tree and so it cannot be part of
the passive tree Therefore the passive acts as a raising verb see section For example
e the tree in Figure would adjoin into a derivation for Bob to talk at the VP to deriv
no de and the mo de passive feature not shown forces the auxiliary to adjoin in as for
other passives as describ ed in chapter
Van exp ects Bob to talk
Bob was exp ected to talk
VPr
V VP*
expected
Figure ECM passive
It has long b een noted that passives of b oth full and bare in nitive ECM constructions are
full in nitives as in and
Bob sees the harmonica fall
The harmonica was seen to fall
The harmonica was seen fall
Under the TAG ECM analysis this fact is easy to implement The fo ot no de of the ECM
passive tree is simply set to have mo de inf whichprevents the derivation of There
fore for all the other trees in the family to fo ot no des are set to have mo de base or
mo de inf dep ending on whether it is a bare in nitiveor not These fo ot no des are all S
no des TheVPfootnodeofthepassive tree however has mo de inf regardless
Sentential Sub jects
Tree families Ts Vnx Ts Ax Ts N Ts Pnx Ts ARBPnx Ts PPnx Ts PNaPnx
Ts V Ts Vtonx Ts NPnx Ts APnx Ts A s
Verbs that select sentential sub jects anchor trees that have an S no de in the sub ject p osition
rather than an NP no de Since extraction is not p ossible from sentential sub jects they are
implemented as substitution no des in the English XTAG grammar Restrictions on sentential
sub jects such as the required that complementizer for indicatives are enforced by feature values
sp eci ed on the S substitution no de in the elementary tree
Sentential sub jects behave essentially like sentential complements with a few exceptions
In general all verbs which license sentential sub jects license the same set of clause typ es Thus
unlikesentential complement verbs whichselect particular complementizers and clause typ es
the matrix verbs licensing sentential sub jects merely license the S argument Information ab out
the complementizer or emb edded verb is lo cated in the tree features rather than in the features
of each verb selecting that tree Thus all sentential sub ject trees have the same mo de
comp and assign comp values shown in Figure a
Sr
Sr NP0↓ VP
V S1* displ-const : set1 : - S0↓ extracted : - VP NA assign-comp : inf_nil/ind_nil inv : - inv : - assign-comp : inf_nil control : <6> punct : contains : <14> comp : that/for/whether/nil comp : that/nil mode : inf/ind mode : ind wh : -
V NP1↓
perplexes thinks
a b
Figure Comparison of assign comp values for sentential sub jects s Vnx a and
sentential complements nx Vs b
The ma jor di erence in clause typ es licensed by S sub js and S comps is that indicativeS
sub js obligatorily have a complementizer see examples in section The assign comp
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
feature is used here to license a null complementizer for in nitival but not indicative clauses
assign comp has the same p ossible values as comp with the exception that the nil
value is split into ind nil and inf nil This di erence in feature values is illustrated in
Figure
Another minor di erence is that whether but not if is grammatical with S sub js Thus if
is not among the comp values allowed in S sub js The nal di erence from S comps is that
there are no S sub js with mo de ger As noted in fo otnote of this chapter gerundive
complements are only allowed when there is no corresp onding NP parse In the case of gerundive
S sub js there is always an NP parse available
Nouns and Prep ositions taking Sentential Complements
Trees NXNs vxPs Pss nxPs Tnx N s Tnx A s
Sr
S1 Sf* NA NA NPr
P◊ S↓ rel-pron : <9> punct : struct : nil N◊ S↓ mode : inf/ind displ-const : set1 : - comp : that/nil inv : - extracted : - inv : -
wh : - extracted : -
a b
Figure Sample trees for prep osition Pss a and noun NXNs b taking sentential
complements
Prep ositions and nouns can also select sentential complements using the trees listed ab ove
These trees use the mo de and comp features as shown in Figure For example the
noun claim takes only indicative complements with that while the prep osition with takes small
clause complements as seen in sentences
Beth s claim that Clovewas a smart dog
Beth s claim that Clove a smart dog
Dania wasn t getting any sleep with Doug sick
Dania wasn t getting any sleep with Doug was sick
Some sp eakers also nd if as a complementizer only marginally grammatical in S comps
PRO control
Typ es of control
In the literature on control two typ es are often distinguished obligatory control as in sen
tences and and optional control as in sentence
Srini promised Mickey PRO to leave
i i i
Srini p ersuaded Mickey PRO to leave
i i
Srini wanted PRO to leave
i i
Christy left the party early PRO to go to the airp ort
i i
PRO to dance is imp ortant for Bill
i
ar b i
At present an analysis for obligatory control into complement clauses as in sentences
and has b een implemented An analysis for cases of obligatory control into adjunct
trol exists and can b e found in Bhatt clauses and optional con
A feature based analysis of PRO control
The analysis for obligatory control involves co indexation of the control feature of the NP
anchored by PRO to the control feature of the controller A feature equation in the tree
anchored by the control verb co indexes the control feature of the controlling NP with the fo ot
no de of the tree All sentential trees have a co indexed control feature from the ro ot Sto the
sub ject NP
When the tree containing the controller adjoins onto the complement clause tree containing
the PRO the features of the fo ot no de of the auxiliary tree are uni ed with the b ottom features
of the ro ot no de of the complement clause tree containing the PRO This leads to the control
feature of the controller b eing co indexed with the control feature of the PRO
Dep ending on the choice of the controlling verb the control propagation paths in the auxil
iary trees are di erent In the case of sub ject control as in sentence the sub ject NP and
the fo ot no de are have co indexed control features while for ob ject control e g sentence
the ob ject NP and the fo ot no de are co indexed for control Among verbs that b elong to the
Tnx Vnx s family i e verbs that take an NP ob ject and a clausal complement sub ject control
verbs form a distinct minority promise b eing the only commonly used verb in this class
Consider the derivation of sentence The auxiliary tree for persuade shown in Figure
has the following feature equation
NP control S t control
The auxiliary tree adjoins into the tree for leave shown in Figure which has the following
feature equation
control S b control NP t
r
Since the adjunction takes place at the ro ot no de S ofthe leave tree after uni cation NP
r
of the persuade tree and NP of the leave tree share a control feature The resulting derived
and derivation trees are shown in Figures and
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
Sr
NP0↓ VP
V NP1↓ control : <1> S2* control : <1> NA
persuaded
Figure Tree for persuaded
Sr control : <1>
NP0↓ control : <1> VP
V
leave
Figure Tree for leave
The nature of the control feature
The control feature do es not have any value and is used only for co indexing purp oses If two
NPs have their control features co indexed it means that they are participating in a relationship
of control the c commanding NP controls the c commanded NP
Long distance transmission of control features
Cases involving emb edded in nitival complements with PRO sub jects such as can also b e
handled
John wants PRO to want PRO to dance
i i i
The control feature of John and the two PRO s all get co indexed This treatment might
app ear to lead to a problem Consider
John wants Mary to want PRO to dance
i i i
e the control feature of their sub ject co indexed to their fo ot If b oth the want trees hav
no des wewould have a situtation where the PRO is co indexed for control feature with John r
NP VP
N V NP control : <1> S2 control : <1> NA
Srini persuaded N NP control : <1> VPr
Mickey PRO V VP NA
to V
leave
Figure Derived tree for Srini persuaded Mickey to leave
αnx0V[leave]
βnx0Vnx1s2[persuaded] (0) αNX[PRO] (1) βVvx[to] (2)
αNXN[Srini] (1) αNXN[Mickey] (2.2)
Figure Derivation tree for Srini persuaded Mickey to leave
as well as with Mary Note that the higher want in is want it assigns case to the
EC M
sub ject of the lower clause while the lower want in is not Sub ject control is restricted to
non ECM Exceptional Case Marking verbs that take in nitival complements Since the two
want s in are di erent with resp ect to their control and other prop erties the control
feature of PRO stops at Mary and is not transmitted to the higher clause
Lo cality constraints on control
PRO control ob eys lo cality constraints The controller for PRO has to b e in the immediately
higher clause Consider the ungrammatical sentence is ungrammatical only with the
co indexing indicated b elow
to p ersuade Mary PRO to dance John wants PRO
i i i i
CHAPTER SENTENTIAL SUBJECTS AND SENTENTIAL COMPLEMENTS
However such a derivation is ruled out automatically by the mechanisms of a TAG derivation
and feature uni cation Supp ose it was p ossible to rst comp ose the want tree with the dance
tree and then insert the persuade tree This is not p ossible in the XTAG grammar b ecause of
the convention that auxiliary trees have NA Null Adjunction constraints on their fo ot no des
Even then at the end of the derivation the control feature of the sub ject of want would end
up co indexed with the PRO sub ject of persuade and the control feature of Mary would be
co indexed with the PRO sub ject of dance as desired There is no way to generate the illegal
co indexing in Thus the lo calit y constraints on PROcontrol fall out from the mechanics
of TAG derivation and feature uni cation
Rep orted sp eech
Rep orted sp eech is handled in the XTAG grammar byhaving the rep orting clause adjoin into
the quote Thus the rep orting clause is an auxiliary tree anchored by the rep orting verb See
Doran for details of the analysis There are trees in b oth the Tnx Vs and Tnx nx s
families to handle rep orting clauses which precede follow and come in the middle of the quote
Chapter
The English Copula Raising Verbs
and Small Clauses
The English copula raising verbs and small clauses are all handled in XTAG by a common
analysis based on sentential clauses headed by non verbal elements Since there are a number of
di erent analyses in the literature of how these phenomena are related or not we will present
rst the data for all three phenomena then various analyses from the literature nishing with
the analysis used in the English XTAG grammar
Usages of the copula raising verbs and small clauses
Copula
The verb be as used in sentences is often referred to as the copula It can b e followed
by a noun adjective or prep ositional phrase
Carl is a jerk
Carl is upset
Carl is in a foul mo o d
Although the copula maylooklike a main verb at rst glance its syntactic b ehavior follows
the auxiliary verbs rather than main verbs In particular
Copula be inv erts with the sub ject
is Beth writing her dissertation
is Beth upset
wrote Beth her dissertation
Copula be o ccurs to the left of the negativemarker not
This chapter is strongly based on Heyco ck Sections and are greatly condensed from her pap er
while the description of the XTAG analysis in section is an up dated and expanded version
CHAPTER THE ENGLISH COPULA RAISING VERBS AND SMALL CLAUSES
Beth is not writing her dissertation
Beth is not upset
Beth wrote not her dissertation
Copula be can contract with the negativemarker not
Beth isn t writing her dissertation
Beth isn t upset
Beth wroten t her dissertation
Copula be can contract with pronominal sub jects
She s writing her dissertation
She s upset
She ote her dissertation
Copula be o ccurs to the left of adverbs in the unmarked order
Beth is often writing her dissertation
Beth is often upset
Beth wrote often her dissertation
Unlike all the other auxiliaries however copula be is not followed bya verbal category by
de nition and therefore must b e the rightmost verb In this resp ect it is likea mainverb
The semantic b ehavior of the copula is also unlikemainverbs In particular any semantic
restrictions or roles placed on the sub ject come from the complement phrase NP AP PP rather
than from the verb as illustrated in sentences and Because the complement phrases
predicate over the sub ject these typ es of sentences are often called predicative sentences
The bartender was garrulous
The cli was garrulous
Raising Verbs
Raising verbs are the class of verbs that share with the copula the prop erty that the complement
rather than the verb places semantic constraints on the sub ject
Carl seems a jerk
Carl seems upset
Carl seems in a foul mo o d
Carl app ears a jerk
Carl app ears upset
Carl app ears in a foul mo o d
The raising verbs are similar to auxiliaries in that they order with other verbs but they
are unique in that they can app ear to the left of the in nitive as seen in the sentences in
They cannot however invert or contract like other auxiliaries and they app ear to the
rightofadverbs
Carl seems to b e a jerk
Carl seems to b e upset
Carl seems to b e in a foul mo o d
seems Carl to b e a jerk
Carl seemn t to b e upset
Carl ems to b e in a foul mo o d
Carl often seems to b e upset
Carl seems often to b e upset
Small Clauses
One way of describing small clauses is as predicativesentences without the copula Since matrix
clauses require tense these clausal structures app ear only as emb edded sentences They o ccur
as complements of certain verbs each of whichmay allow certain typ es of small clauses but not
others dep ending on its lexical idiosyncrasies
I consider Carl a jerk
I consider Carl upset
I consider Carl in a foul mo o d
I prefer Carl in a foul mo o d
I prefer Carl upset
Raising Adjectives
Raising adjectives are the class of adjectives that share with the copula and raising verbs the
prop erty that the complement rather than the verb places semantic constraints on the sub ject
They app ear with the copula in a matrix clause as in However in other cases such
as that of small clauses they do not have to app ear with the copula
Carl is likelytobeajerk
Carl is likely to b e upset
Carl is likely to b e in a foul mo o d
Carl is likely to p erjure himself
I consider Carl likely to p erjure himself
Various Analyses
Main Verb Raising to INFL Small Clause
In Pollo ck the copula is generated as the head of a VP like any main verb such as
sing or buy Unlike all other main verbs however be moves out of the VP and into In in a
with the exception of have in British English See fo otnote in Chapter
CHAPTER THE ENGLISH COPULA RAISING VERBS AND SMALL CLAUSES
tensed sentence This analysis aims to account for the b ehavior of be as an auxiliary in terms
of inversion negative placement and adverb placement while retaining a sentential structure
in which be heads the main VP at D Structure and can thus b e the only verb in the clause
Pollo ck claims that the predicative phrase is not an argumentofbe which instead he assumes
to take a small clause complement consisting of a no de dominating an NP and a predicative
AP NP or PP The sub ject NP of the small clause then raises to b ecome the sub ject of the
sentence This accounts for the failure of the copula to imp ose any selectional restrictions on
the sub ject Raising verbs such as seem and appear presumably take the same typ e of small
clause complement
Auxiliary Null Copula
In Lap ointe the copula is treated as an auxiliary verb that takes as its complementaVP
headed by a passiveverb a present participle or a null verb the true copula This verb may
then take AP NP or PP complements The author points out that there are many languages
that have b een analyzed as having a null copula but that English has the p eculiaritythat its
null copula requires the co presence of the auxiliary be
Auxiliary Predicative Phrase
In GPSG Gazdar et al Sag et al the copula is treated as an auxiliary verb
that takes an X category with a value for the head feature PRD predicative AP NP PP
and VP can all b e PRD but a Feature Co o ccurrence Restriction guarantees that a PRD
VP will b e headed bya verb that is either passive or a present participle
GPSG follows Chomsky in adopting the binary valued features V and N for de
comp osing the verb noun adjective and prep osition categories In that analysis verbs are
V N nouns are V N adjectives V N and prep ositions V N NP and AP
predicative complements generally pattern together a fact that can be stated economically
using this category decomp osition In neither Sag et al nor Chomsky is there
any discussion of how to handle the complete range of complements to a verb like seem which
takes AP NP and PP complements as well as in nitives The solution would app ear to b e to
asso ciate the verb with two sets of rules for small clauses leaving aside the use of the verb with
an expletive sub ject and sentential complement
Auxiliary Small Clause
In Moro the copula is treated as a sp ecial functional category a lexicalization of tense
which is considered to head its own pro jection It takes as a complement the pro jection of
another functional category Agr agreement This pro jection corresp onds roughly to a small
clause and is considered to b e the domain within which predication takes place An NP must
then raise out of this pro jection to become the sub ject of the sentence it may b e the sub ject
of the AgrP or if the predicate of the AgrP is an NP this ma y raise instead In addition
to o ccurring as the complement of be AgrP is selected by certain verbs such as consider It
follows from this analysis that when the complementto consider is a simple AgrP it will always
consist of a sub ject followed by a predicate whereas if the complement contains the verb be
the predicate of the AgrP may raise to the left of be leaving the sub ject of the AgrP to the
right
John is t the culprit
i Ag r P i
The culprit is John t
i Ag r P i
I consider John the culprit
Ag r P
I consider John to b e t the culprit
i Ag r P i
I consider the culprit to b e John t
i AgrP i
Moro do es not discuss a numb er of asp ects of his analysis including the nature of Agr and
the implied existence of sentences without VP s
XTAG analysis
Sr Sr Sr
NP ↓ VP NP ↓ VP NP0↓ VP 0 0
V NP1 V AP1 V PP1
ε N◊ ε A◊ ε P◊ NP1↓
a b c
Figure Predicative trees nx N a nx Ax b and nx Pnx c
The XTAG grammar provides a uniform analysis for the copula raising verbs and small
clauses by treating the maximal pro jections of lexical items that can b e predicated as predicative
clauses rather than simply noun adjective and prep ositional phrases The copula adjoins in
for matrix clauses as do the raising verbs Certain other verbs suchas consider can take the
predicative clause as a complement without the adjunction of the copula to form the emb edded
small clause
The structure of a predicative clause then is roughly as seen in for NP s AP s
are shown in Figures a and PP s The XTAG trees corresp onding to these structures
b and c resp ectively
There are actually two other predicative trees in the XTAG grammar Another predicative noun phrase tree
is needed for noun phrases without determiners as in the sentence They are remen and another prep ositional
phrase tree is needed for exhaustive prep ositional phrases suchas The workers arebelow
CHAPTER THE ENGLISH COPULA RAISING VERBS AND SMALL CLAUSES
NP N
S VP
NP A
S VP
NP P
S VP
The copula be and raising verbs all get the basic auxiliary tree as explained in the section
on auxiliary verbs section Unlike the raising verbs the copula also selects the inverted
auxiliary tree set Figure shows the basic auxiliary tree anchored by the copula be The
mo de feature is used to distinguish the predicative constructions so that only the copula
and raising verbs adjoin onto the predicative trees
VPr conditional : <9> perfect : <10> progressive : <11> displ-const : set1 : <8> assign-case : <7> mode : <6> tense : <5> agr : <3> neg : <2> assign-comp : <1> mainv : <4>
V assign-comp : <1> VP* displ-const : set1 : - neg : <2> NA progressive : <11> agr : <3> perfect : <10> mainv : <4> - conditional : <9> tense : <5> mode : nom/prep mode : <6> assign-case : <7> displ-const : set1 : <8> mode : ind tense : pres mainv : - assign-comp : ind_nil/adj/that/rel/if/whether assign-case : nom agr : 3rdsing : + num : sing pers : 3
is
Figure Copula auxiliary tree Vvx
There are two p ossible values of mo de that corresp ond to the predicative trees nom
and prep They corresp ond to a mo di ed version of the four valued N V feature describ ed
in section The nom value corresp onds to N selecting the NP and AP predicative
clauses As mentioned earlier they often pattern together with resp ect to constructions using
predicative clauses The remaining prep ositional phrase predicative clauses then corresp ond
to the prep mo de
Figure shows the predicative adjective tree from Figure b nowanchored by upset and
with the features visible As mentioned mo de nom on the VP no de prevents auxiliaries
other than the copula or raising verbs from adjoining into this tree In addition it prevents the
predicative tree from o ccurring as a matrix clause Since all matrix clauses in XTAGmust be
mo de indicative ind or imp erative imp a tree with mo de nom or mo de prep
must have an auxiliary verb the copula or a raising verb adjoin in to makeit mo de ind
The distribution of small clauses as emb edded complements to some verbs is also man
aged through the mo de feature Verbs such as consider and prefer select trees that take
a sentential complement and then restrict that complement to be mo de nom and or
mo de prep dep ending on the lexical idiosyncrasies of that particular verb Many verbs
that don t take small clause complements do takesentential complements that are mo de ind
which includes small clauses with the copula already adjoined Hence as seen in sentence sets
consider takes only small clause complements prefer takes b oth prep but not nom
small clauses and indicative clauses while fe el takes only indicative clauses
She considers Carl a jerk
She considers Carl in a foul mo o d
She considers that Carl is a jerk
She prefers Carl a jerk
She prefers Carl in a foul mo o d
She prefers that Carl is a jerk
She feels Carl a jerk
She feels Carl in a foul mo o d
She feels that Carl is a jerk
Figure shows the tree anchored by consider that takes the predicative small clauses
Raising verbs such as seems work essentially the same as the auxiliaries in that they also
select the basic auxiliary tree as in Figure The only di erence is that the value of mo de
on the VP fo ot no de mightbedi erent dep ending on what typ es of complements the raising
verb takes Also two of the raising verbs take an additional tree Vpxvx shown in Figure
which allows for an exp eriencer argument as in John seems to me to behappy
Raising adjectives such as likely take the tree shown in Figure This tree combines
asp ects of the auxiliary tree Vvx and the adjectival predicative tree shown in Figure b
As with Vvx it adjoins in as a VP auxiliary tree However since it is anchored by an adjective
not a verb it is similar to the adjectival predicative tree in that it has an at the V no de and
a feature value of mo de nom which is passed up to the VP ro ot indicates that it is an
adjectival predication This serves the same purp ose as in the case of the tree in Figure
and forces another auxiliary verb such as the copula to adjoin in to makeit mo de ind
CHAPTER THE ENGLISH COPULA RAISING VERBS AND SMALL CLAUSES
Sr displ-const : set1 : - assign-comp : inf_nil/ind_nil assign-case : <1> agr : <2> tense : <3> comp : nil mainv : <4> mode : <5> displ-const : set1 : <6> assign-comp : <7> inv : - extracted : -
NP0↓ wh : - VP assign-case : <1> case : <1> agr : <2> agr : <2> tense : <3> mainv : <4> mode : <5> displ-const : set1 : <6> assign-comp : <7> assign-case : acc mode : nom displ-const : set1 : -
V AP1
ε A wh : -
upset
Figure Predicative AP tree with features nx Ax
Non predicative BE
The examples with the copula that we have given seem to indicate that be is always followed
by a predicative phrase of some sort This is not the case however as seen in sentences suchas
The noun phrases in these sentences are not predicative They do not take raising
verbs and they do not o ccur in emb edded small clause constructions
my teacher is Mrs Wayman
Doug is the man with the glasses Sr
NP0↓ VP
V S1* displ-const : set1 : - assign-comp : inf_nil/ind_nil inv : - assign-case : acc comp : nil mode : nom/prep
consider
Figure Consider tree for emb edded small clauses
VPr
V◊ PP VP* NA
P NP↓
to
Figure Raising verb with exp eriencer tree Vpxvx
VPr
V AP NA
ε A◊ VP*
NA
Figure Raising adjective tree Vvx adj
My teacher seems Mrs Wayman
Doug app ears the man with the glasses
CHAPTER THE ENGLISH COPULA RAISING VERBS AND SMALL CLAUSES
I consider my teacher Mrs Wayman
I prefer Doug the man with the glasses
In addition the sub ject and complementcan exchange p ositions in these typ e of examples
but not in sentences with predicative be Sentence has the same interpretation as sentence
and di ers only in the p ositions of the sub ject and complement NP s Similar sentences
with a predicative be are shown in and In this case the sentence with the exchanged
NP s is ungrammatical
The man with the glasses is Doug
Doug is a programmer
A programmer is Doug
The non predicative be in and also called equative be patterns di erently both
syntactically and semantically from the predicative usage of be Since these sentences are
clearly not predicative it is not desirable to have a tree structure that is anchored bytheNP
AP orPP asweha ve in the predicativesentences In addition to the conceptual problem we
would also need a mechanism to blo ck raising verbs from adjoining into these sentences while
allowing them for true predicative phrases and prevent these typ es of sentence from b eing
emb edded again while allowing them for true predicative phrases
Sq
V Sr Sr
is NP0↓ VPr
NP0↓ VPr NA
V VP1 Vr VP1
ε ↓ is V1 NP1↓ V1 NP1
ε1 ε1
a b
Figure Equative BE trees nx BEnx a and Invnx BEnx b
Although non predicative be is not a raising verb it do es exhibit the auxiliary verb b ehavior
set out in section It inverts contracts and so forth as seen in sentences and
and therefore can not be asso ciated with any existing tree family for main verbs It requires
a separate tree family that includes the tree for inversion Figures a and b show the
declarativeandinverted trees resp ectively for equative be
is my teacher Mrs Wayman
Doug isn t the man with the glasses
Chapter
Ditransitive constructions and
dative shift
Verbs such as give and put that require two ob jects as shown in examples are
termed ditransitive
Christygave a cannoli to Beth Ann
Christygave Beth Ann
Christy put a cannoli in the refrigerator
Christy put a cannoli
The indirect ob jects Beth Ann and refrigerator app ear in these examples in the form of
PP s Within the set of ditransitiveverbs there is a subset that also allowtwo NP s as in
As can b e seen from and this twoNP or double ob ject construction is grammatical
for give but not for put
Christygave Beth Ann a cannoli
Christy put the refrigerator the cannoli
The alternation b etween and is known as dative shift In order to account for
verbs with dative shift the English XTAG grammar includes structures for b oth variants in the
tree family Tnx Vnx Pnx The declarative trees for the shifted and non shifted alternations
are shown in Figure
The indexing of no des in these two trees represents the fact that the semantic role of
the indirect ob ject NP in Figure a is the same as that of the direct ob ject NP in
Figure b and vice versa This use of indexing is consistent with our treatment of other
constructions suchaspassive and ergative
Verbs that do not show this alternation and have only the NP PP structure e g put select
the tree family Tnx Vnx pnx Unlike the Tnx Vnx Pnx family the Tnx Vnx pnx tree
family do es not contain trees for the NP NP structure Other verbs suchas ask allow only the
NP NP structure as shown in and
In languages similar to English that haveovert case marking indirect ob jects would b e marked with dative
case It has also b een suggested that for English the prep osition to serves as a dative case marker
CHAPTER DITRANSITIVE CONSTRUCTIONS AND DATIVE SHIFT
S r
NP 0↓ VP
S V◊ NP 1↓ PP 2 r
↓ P2 NP 2 NP 0↓ VP
to V◊ NP 2↓ NP 1↓
a b
Figure Dative shift trees nx Vnx Pnx a and nx Vnx nx b
Beth Ann asked Srini a question
Beth Ann asked a question to Srini
Verbs that only allow the NP NP structure select the tree family Tnx Vnx nx This tree
family do es not have the trees for the NP PP structure
Notice that in Figure a the prep osition to is built into the tree There are other
apparent cases of dative shift with for such as in and that we have taken to be
structurally distinct from the cases with to
Beth Ann baked Dusty a biscuit
Beth Ann baked a biscuit for Dusty
McCawley notes
A for dative expression in underlying structure is external to the V with whichit
is combined in view of the fact that the latter b ehaves as a unit with regard to all
relevant syntactic phenomena
In other words the for PP s that app ear to undergo dative shift are actually adjuncts not
complements Examples and demonstrate that PP s with for are optional while
ditransitive to PP s are not
Beth Ann made dinner
Beth Ann gave dinner
Consequently intheXTAG grammar the apparent dative shift with for PP s is treated as
Tnx Vnx nx for the NP NP structure and as a transitive plus an adjoined adjunct PP for the
NP PP structure To account for the ditransitive to PP s the prep osition to is built into the
tree family Tnx Vnx tonx This accounts for the fact that to is the only prep osition allowed
in dative shift constructions
McCawley also notes that the to and for cases di er with resp ect to passivization the
indirect ob jects with to may b e the sub jects of corresp onding passives while the alleged indirect
ob jects with for cannot as in sentences Note that the passivisation examples are
for NP NP structures of verbs that take to or for PP s
Beth Ann gave Clove dinner
Clovewas given dinner by Beth Ann
Beth Ann made Clove dinner
Clovewas made dinner by Beth Ann
However we b elieve that this to b e incorrect and that the indirect ob jects in the for case
are allowed to b e the sub jects of passives as in sentences The apparent strangeness
of sentence is caused byinterference from other interpretations of Clove was made dinner
Dania baked Doug a cake
Doug was baked a cakeby Dania
Chapter
It clefts
There are several varieties of it clefts in English All the it clefts have four ma jor comp onents
the dummy sub ject it
the main verb be
the clefted element A constituent XP compatible with any gap in the clause
the clause A clause e g S with or without a gap
Examples of it clefts are shown in
it was here that the ENIACwas created
XP XP S S
it was at MIT that colorless green ideas slept furiously
XP XP S S
it is happily that Seth quit Reality
XP XP S S
it was there that she would have to enact her renunciation
XP XP S S
The clefted elementcanbeofanumb er of categories for example NP PPoradverb The
clause can also be of several typ es The English XTAG grammar currently has a separate
analysis for only a subset of the sp eci cational it clefts in particular the ones without gaps
in the clause e g and It clefts that have gaps in the clause such as and
are currently handled as relative clauses Although arguments have b een made against
treating the clefted element and the clause as a constituent Delahunty the relative
clause approach do es capture the restriction that the clefted element must ll the gap in the
clause and do es not require any additional trees
In the sp eci cational it cleft without gaps in the clause the clefted element has the role
of an adjunct with resp ect to the clause For these cases the English XTAG grammar requires
additional trees These it cleft trees are in separate tree families b ecause although some re
searchers e g Akma jian derived it clefts through movement from other sentence typ es
most current researchers e g Delahunty Knowles Gazdar et al Delin
See e g Ball Delin and Delahunty for more detailed discussion of typ es of it clefts
and Sornicola favor base generation of the various cleft sentences Placing the
it cleft trees in their own tree families is consistent with the current preference for base genera
tion since in the XTAG English grammar structures that would b e related by transformation
in a movement based account will app ear in the same tree family Like the base generated
approaches the placement of it clefts in separate tree families makes the claim that there is no
derivational relation b etween it clefts and other sentence typ es
The three it cleft tree families are virtually identical except for the category lab el of the
clefted element Figure shows the declarative tree and an inverted tree for the PP It cleft
tree family
S q
S r V◊ S r NA
NP 0 VP NP 0 VPr
N◊ V◊ VP1 N◊ Vr VP1
V1 PP 1↓ S 2↓ ε V1 PP 1↓ S 2↓
ε ε1
a b
Figure It cleft with PP clefted element ItVpnx s a and InvItVpnx s b
The extra layer of tree structure in the VP represents that while be is a main verb rather
than an auxiliary in these cases it retains some auxiliary prop erties The VP structure for the
equative it cleft be is identical to that obtained after adjunction of predicative be into small
clauses The inverted tree in Figure b is necessary b ecause of be s auxiliary likebehavior
Although be is the main verb in it clefts it inverts like an auxiliary Main verb inversion cannot
b e accomplished by adjunction as is done with auxiliaries and therefore must b e built into the
tree family The tree in Figure b is used for yes no questions suchas
was it in the forest that the wolf talked to the little girl
For additional discussion of equative or predicative be see Chapter
CHAPTER IT CLEFTS
Part III
Sentence Typ es
Chapter
Passives
In passive constructions such as the sub ject NP is interpreted as having the same role
as the direct ob ject NP in the corresp onding active declarative
An airline buy out bill was approved by the House WSJ
The House approved an airline buy out bill
S r S r S r mode : <3>
↓ ↓ NP 1 VP NP 1 VP
NP1↓ VP mode : <3> passive : <1> V ◊ PP S * V ◊ S * PP mode : <2> 2 2
↓ ↓ P NP 0 P NP 0 V◊ passive : <1> + S2* mode : <2> ppart
by by
a b c
Figure Passive trees in the Sentential Complement with NP tree family nx Vs a
nx Vbynx s b and nx Vs bynx c
In a movement analysis the direct ob ject is said to havemoved to the sub ject p osition The
original declarative sub ject is either absent in the passiveorisina by headed PP by phrase
In the English XTAG grammar passive constructions are handled by having separate trees
within the appropriate tree families Passive trees are found in most tree families that have a
direct ob ject in the declarative tree the light verb tree families for instance do not contain
e trees o ccur in pairs one tree with the by phrase and another without passive trees Passiv
CHAPTER PASSIVES
it Variations in the lo cation of the by phrase are p ossible if a sub categorization includes other
arguments such as a PP or an indirect ob ject Additional trees are required for these variations
For example the Sentential Complement with NP tree family has three passive trees shown
in Figure one without the by phrase Figure a one with the by phrase before
the sentential complement Figure b and one with the by phrase after the sentential
complement Figure c
Figure a also shows the feature restrictions imp osed on the anchor Only verbs with
mo de ppart i e verbs with passive morphology can anchor this tree The mo de
feature is also resp onsible for requiring that passive be adjoin into the tree to create a matrix sen
tence Since a requirement is imp osed that all matrix sentences must have mo de ind imp
an auxiliary verb that selects mo de ppart and passive suchaswas must adjoin
see Chapter for more information on the auxiliary verb system
A reduced set of features are shown for readability
Chapter
Extraction
The discussion in this chapter covers constructions that are analyzed as having wh movement
in GB in particular wh questions and topicalization Relative clauses which could also be
considered extractions are discussed in Chapter
Extraction involves a constituent app earing in a linear p osition to the left of the clause with
whichitisinterpreted One clause argument p osition is empty For example the p osition lled
by frisbee in the declarativeinsentence is emptyinsentence The wh item what in
sentence is of the same syntactic category as frisbee in sentence and lls the same
role with resp ect to the sub categorization
Clove caught a frisb ee
What did Clove catch
i i
The English XTAG grammar represents the connection b etween the extracted elementand
the empty p osition with co indexing as do es GB The trace feature is used to implement
the co indexing In extraction trees in XTAG the empty p osition is lled with an The
extracted item always app ears in these trees as a sister to the S tree with b oth dominated by
r
aS ro ot no de The S subtrees in extraction trees have the same structure as the declarative
q r
tree in the same tree family The additional structure in extraction trees of the S and NP
q
no des roughly corresp onds to the CP and Sp ec of CP p ositions in GB
All sentential trees with extracted comp onents thisdoesnot include relative clause trees
are marked extracted at the top S no de while sentential trees with no extracted
comp onents are marked extracted Items that takeemb edded sentences such as nouns
verbs and some prep ositions can place restrictions on whether the em b edded sentence is allowed
to b e extracted or not For instance sentential sub jects and sentential complements of nouns
and prep ositions are not allowed to be extracted while certain verbs may allow extracted
sentential complements and others may not e g sentences
The jury wondered who killed Nicole
The jury wondered who Simpson killed
The jury thought Simpson killed Nicole
CHAPTER EXTRACTION
The jury thought who did Simpson kill
The extracted feature is also used to blo ckemb edded topicalization in in nitival comple
ment clauses as exempli ed in
John w ants Bill PROtoseet
i i
Verbs such as want that take non wh in nitival complements sp ecify that the extracted
feature of their complement clause i e of the fo ot Snode is Clauses that involve topical
ization have as the value of their extracted feature i e of the ro ot S no de Sentences
like are thus ruled out
Sq invlink : <1> inv : <1> extracted : + wh : <5>
NP↓ case : <2> Sr inv : <1> agr : <3> inv : - trace : <4> wh : <5>
NP0↓ VP
V◊ NP1 case : acc NA case : <2> agr : <3> trace : <4>
ε
Figure Transitive tree with ob ject extraction W nx Vnx
The tree that is used to derive the embedded sentence in in the English XTAG
grammar is shown in Figure The imp ortant features of extracted trees are
The subtree that has S as its ro ot is identical to the declarative tree or a non extracted
r
passive tree except for having one NP p osition in the VP lled by
The ro ot S no de is S which dominates NP and S
q r
The trace feature of the lled NP is co indexed with the trace feature of the
NP daughter of S
q
Features not p ertaining to this discussion have b een taken out to improve readability
The case and agr features are passed from the empty NP to the extracted NP This
is particularly imp ortant for extractions from sub ject NP s since case can continue
to b e assigned from the verb to the sub ject NP p osition and from there b e passed to the
extracted NP
The inv feature of S is co indexed to the wh feature of NP through the use of
r
the invlink feature in order to force sub ject auxiliary inversion where needed see
section for more discussion of the inv wh co indexing and the use of these
trees for topicalization
Topicalization and the v alue of the inv feature
Our analysis of topicalization uses the same trees as wh extraction For any NP complement
position a single tree is used for b oth wh questions and for topicalization from that p osition
Wh questions have sub ject auxiliary inversion and topicalizations do not This di erence be
tween the constructions is captured by equating the values of the S s inv feature and the
r
extracted NP s wh feature This means that if the extracted item is a wh expression as in
wh questions the value of inv will b e andaninverted auxiliary will b e forced to adjoin
If the extracted item is a non wh inv will b e and no auxiliary adjunction will o ccur An
additional complication is that inversion only o ccurs in matrix clauses so the values of inv
and wh should only be equated in matrix clauses and not in embedded clauses In the
English XTAG grammar appropriate equating of the inv and wh features is accom
plished using the invlink feature and a restriction imp osed on the ro ot S of a derivation In
particular in extraction trees that are used for b oth wh questions and topicalizations the value
of the inv feature for the top of the S no de is co indexed to the value of the inv feature
r
on the b ottom of the S no de On the b ottom of the S no de the inv feature is co indexed
q q
to the invlink feature The wh feature of the extracted NP no de is co indexed to the
value of the wh feature on the b ottom of S The linking b etween the value of the S wh
q q
and the invlink features is imp osed by a condition on the nal ro ot no de of a derivation
i e the top S no de of a matrix clause requires that invlink wh For example the
tree in Figure is used to derive b oth and
John I like
Who do you like
For the question in the extracted item who has the feature value wh sothe
value of the inv feature on VP is also and an auxiliary in this case do is forced to adjoin
feature and for S s inv feature For the topicalization the values for John s wh
q
are b oth and no auxiliary adjoins
Extracted sub jects
The extracted sub ject trees provide for sentences like dep ending on the tree family
with which it is asso ciated
CHAPTER EXTRACTION
Who left
Who wrote the pap er
Who was happy
Wh questions on sub jects di er from other argument extractions in not having sub ject
auxiliary inversion This means that in sub ject wh questions the linear order of the constituents
is the same as in declaratives so it is di cult to tell whether the sub ject has moved out of p osition
or not see Heyco ck and Kro ch for arguments for and against moved sub ject
The English XTAG treatment of sub ject extractions assumes the following
Syntactic sub ject topicalizations don t exist and
Sub jects in wh questions are extracted rather than in situ
The assumption that there is no syntactic sub ject topicalization is reasonable in English
since there is no convincing syntactic evidence and since the interpretability of sub jects as
topics seems to b e mainly a ected by discourse and intonational factors rather than syntactic
structure As for the assumption that wh question sub jects are extracted these questions seem
to have more similarities to other extractions than to the two cases in English that have b een
considered in situ wh multiple wh questions and echo questions In multiple wh questions such
as sentence one of the wh items is blo cked from moving sentence initially b ecause the
rst wh item already o ccupies the lo cation to whichitwould move
Who ate what
This typ e of blo cking account is not applicable to sub ject wh questions b ecause there is
no obvious candidate to do the blo cking Similarity between sub ject wh questions and echo
questions is also lacking At least one account of echo questions Ho ckey argues that
ec ho questions are not ordinary wh questions at all but rather fo cus constructions in which the
wh item is the fo cus Clearly this is not applicable to sub ject wh questions So it seems that
treating sub ject wh questions similarly to other wh extractions is more justi ed than an in situ
treatment
Given these assumptions there must b e separate trees in each tree family for sub ject extrac
tions The declarative tree cannot b e used even though the linear order is the same b ecause the
structure is di erent Since topicalizations are not allowed the wh feature for the extracted
NP no de is set in these trees to The lack of sub ject auxiliary inversion is handled by the
absence of the invlink feature Without the presence of this feature the wh and inv
are never linked so inversion can not o ccur Like other wh extractions the S no de is marked
q
extracted to constrain the o ccurrence of these trees in emb edded sentences The tree
in Figure is an example of a sub ject wh question tree
Wh moved NP complement
Wh questions can b e formed on every NP ob ject or indirect ob ject that app ears in the declar
passive trees as seen in sentences A tree family will contain ative tree or in the Sq inv : <9> wh : <8> extracted : +
NP↓ agr : <5> Sr inv : <9> case : <6> wh : <8> trace : <7> assign-case : <3> wh : <8> + agr : <4> inv : -
NP0 case : <3> VP assign-case : <3> NA agr : <4> agr : <4> agr : <5> agr : <1> case : <6> assign-case : <2> trace : <7>
ε V◊ agr : <1>
assign-case : <2>
Figure Intransitive tree with sub ject extraction W nx V
one tree for each of these p ossible NP complement p ositions Figure shows the two extrac
tion trees from the ditransitive tree family for the extraction of the direct Figure a and
indirect ob ject Figure b
Dania asked Beth a question
Who did Dania ask a question
i i
What did Dania ask Beth
i i
Beth was asked a question by Dania
Who was Beth asked a question by
i i
What was Beth asked by Dania
i i
Wh moved ob ject of a P
Wh questions can b e formed on the NP ob ject of a complementPPasinsentence
Which dog did Beth Ann give a b one to
i i
CHAPTER EXTRACTION
Sq Sq
NP↓ Sr NP↓ Sr
NP0↓ VP NP0↓ VP
V◊ NP1↓ NP2 V◊ NP1 NP2↓ NA NA
ε ε
a b
Figure Ditransitive trees with direct ob ject W nx Vnx nx a and indirect ob ject
extraction W nx Vnx nx b
The by phrases of passives behave like complements and can undergo the same typ e of
extraction as in
Which dog was the frisb ee caughtby
i i
Tree structures for this typ e of sentence are very similar to those for the wh extraction of NP
complements discussed in section and have the identical imp ortant features related to tree
structure and trace and inversion features The tree in Figure is an example of this typ e
of tree Topicalization of NP ob jects of prep ositions is handled the same way as topicalization
of complement NP s
Sq
NP↓ Sr
NP0↓ VP
V◊ NP1↓ PP2
P2↓ NP2 NA
ε
Figure Ditransitive with PP tree with the ob ject of the PP extracted W nx Vnx pnx
Wh moved PP
Like NP complements PP complements can b e extracted to form wh questions as in sentence
Towhichdog did Beth Ann throw the frisb ee
i i
As can b e seen in the tree in Figure extraction of PP complements is very similar to
extraction of NP complements from the same p ositions
Sq
PP2↓ Sr
NP0↓ VP
V◊ NP1↓ PP NA
ε
Figure Ditransitive with PP with PP extraction tree pW nx Vnx pnx
The PP extraction trees di er from NP extraction trees in having a PP rather than an NP
left daughter no de under S and in having the ll a PP rather than an NP p osition in the
q
VP In other resp ects these PP extraction structures b ehavelike the NP extractions including
b eing used for topicalization
Wh moved S complement
Except for the no de lab el on the extracted p osition the trees for wh questions on S complements
lo ok exactly like the trees for wh questions on NP s in the same p ositions This is b ecause there
is no separate wh lexical item for clauses in English so the item what is ambiguous between
representing a clause or an NP To illustrate this ambiguity notice that the question in
an NP as in The extracted NP in could be answered by either a clause as in or
these trees is constrained to b e wh since sentential complements can not b e topicalized
What do es Clovewant
for Beth Ann to play frisb ee with her
a biscuit
CHAPTER EXTRACTION
Wh moved Adjective complement
In sub categorizations that select an adjective complement that complement can b e questioned
in a wh question as in sentence
How did he feel
i i
Sq
AP↓ Sr
NP0 ↓ VP
V◊ AP1 NA
ε
Figure Predicative Adjective tree with extracted adjective WA nx Vax
The tree families with adjective complements include trees for such adjective extractions
that are very similar to the wh extraction trees for other categories of complements The
adjective p osition in the VP is lled byan and the trace feature of the adjective complement
and of the adjective daughter of S are co indexed The extracted adjective is required to be
q
wh so no topicalizations are allowed An example of this typ e of tree is shown in
Figure
How is the only wh adjective currently in the XTAG English grammar
Chapter
Relative Clauses
Relative clauses are NP mo di ers whichinvolve extraction of an argument or an adjunct The
NP head the p ortion of the NP b eing mo di ed by the relative clause is not directly related to
the extracted element For example in the person is the head NP and is mo di ed by the
relative clause whose mother likes Chris The person is not interpreted as the sub ject of the
relative clause which is missing an overt sub ject In other cases such as the relationship
between the head NP export exhibitions may seem to b e more direct but even there we assume
that there are two indep endent relationships one b etween the entire relative clause and the NP
it mo di es and another between the extracted element and its trace The extracted element
maybeanovert wh phrase as in or a covert element as in
the p erson whose mother likes Chris
exp ort exhibitions that included high techitems
Relative clauses are represented in the English XT AG grammar by auxiliary trees that adjoin
to NP s These trees are anchored by the verb in the clause and app ear in the appropriate tree
families for the various verb sub categorizations Within a tree family there will b e groups of
relative clause trees based on the declarative tree and each passive tree Within each of these
groups there is a separate relative clause tree corresp onding to each p ossible argument that can
b e extracted from the clause There is no relationship b etween the extracted p osition and the
head NP The relationship b etween the relative clause and the head NP is treated as a semantic
relationship which will b e provided by any reasonable comp ositional theory The relationship
between the extracted element whichcan be covert is captured by co indexing the trace
features of the extracted NP and the NP no de in the relative clause tree If for example it is
w
NP that is extracted wehave the following feature equations
NP t h trace i NP t h trace i
w
NP t h case i NP t h case i
w
NP t h agr i NP t h agr i
w
Representative examples from the transitive tree family are shown with a relevant subset
of their features in Figures a and b Figure a involves a relative clause with a
covert extracted element while gure b involves a relative clause with an overt wh phrase
No adjunct traces are represented in the XTAG analysis of adjunct extraction Relative clauses on adjuncts
do not have traces and consequently feature equations of the kind shown here are not present
The convention followed in naming relative clause trees is outlined in App endix D
CHAPTER RELATIVE CLAUSES
NPr NPr
NPf * agr : <12> Sp NPf * case : <10> Sp NA case : <13> NA NA agr : <11> NA wh : <14> wh : <12> case : nom/acc case : nom/acc
NPw ↓ Sq NPw agr : <1> Sq select-mode : <1> ind wh : + NA case : <2> NA NA trace : <3> agr : <2> case : <3> trace : <4>
Comp Sr εw Comp↓ rel-pron : <4> Sr select-mode : <5> inf/ind NA assign-case : <6>
εc NP0 case : <8> VP NA agr : <9> NP0 ↓ VP agr : <2> case : <3> trace : <4> V NP1 case : acc NA agr : <1> case : <2> trace : <3> ε V NP1 ↓
likes ε likes
a b
Figure Relative clause trees in the transitive tree family Nc nx Vnx a and
N nx Vnx b
The ab ove analysis is essentially identical to the GB analysis of relative clauses One asp ect
of its implementation is that an covert wh NP and a covert Comp havetobeintro duced
See and for example
exp ort exhibitions that included high tech items
NP i i
w
the exp ort exhibition Muriel planned
NP i C i
w
The lexicalized nature of XTAG makes it problematic to have trees headed bynull strings
Of the two null trees NP and Comp that we could p ostulate the former is de nitely more
w
undesirable b ecause it would lead to massiveovergeneration as can b e seen in and
did John eat the apple as a wh question
NP
w
Iwonder Mary likes John as an indirect question
NP
w
The presence of an initial headed byanull Comp do es not lead to problems of overgeneration
b ecause relative clauses are the only environment with a Comp substitution no de
Complementizers in clausal complementation are intro duced by adjunction See section
Consequently our treatment of relative clauses has di erent trees to handle relative clauses
with an overt extracted wh NP and relative clauses with a covert extracted wh NP Relative
clauses with an overt extracted wh NP involve substitution of a wh NP into the NP
w
no de and have a Comp no de headed by built in Relative clauses with a covert extracted
C
wh NP have a NP no de headed by built in and involve substitution into the Comp no de
w w
The Comp no de that is intro duced by substitution can b e the n ull complementizer that
C
and for
For example the tree shown in Figure b is used for the relative clauses shown in
sentences while the tree shown in Figure a is used for the relative clauses in
sentences
the man who Muriel likes
the man whose mother Muriel likes
the man Muriel likes
the b o ok for Muriel to read
the man that Muriel likes
the b o ok Muriel is reading
Cases of PP pied piping cf are handled in a similar fashion by building in a PP
w
no de
the demon by whom Muriel was chased
See the tree in Figure
Complementizers and clauses
The co o ccurrence constraints that exist between various Comps and the clause typ e of the
clause they o ccur with are implemented through combinations of di erent clause typ es using
the mo de feature the select mo de feature and the rel pron feature
Clauses are sp eci ed for the mo de feature which indicates the clause typ e of that clause
Possible values for the mo de feature are ind inf ppart ger etc
Comps are lexically sp eci ed for a feature named select mo de In addition the
select mo de feature of the Comp is equated with the mo de feature of its comple
mentSby the following equation
S t hmo dei Comp t hselect mo dei
r
The lexical sp eci cations of the Comps are shown b elow
Comp t hselect mo dei ind inf ger ppart
C
The feature equation used is NP t wh Examples of NPs that could substitute under NP are
w w
whose mother who whom and also which but not when and where which are treated as exhaustive wh PPs
CHAPTER RELATIVE CLAUSES
NPr
NPf * case : <1> Sp NA agr : <2> NA wh : <3> case : nom/acc
PPw ↓ wh : + Sq select-mode : <4> ind/inf NA
Comp Sr NA
εc NP0 ↓ VP
V NP1 ↓
chased
Figure Adjunct relative clause tree with PP pied piping in the transitive tree family
Npxnx Vnx
that Comp t hselect mo dei ind
for Comp t hselect mo dei inf
The following examples display the co o ccurence constraints which the select mo de
sp eci cations assigned ab ove implement
For
C
the b o ok Muriel likes S t mo de ind
a book to like S t mo de inf
the girl reading the b o ok S t mo de ger
the book read by Muriel S t mo de ppart
or for F
the b o ok for Muriel likes S t mo de ind
a b o ok for Mary to like S t mo de inf
the girl for reading the b o ok S t mo de ger
the b o ok for read byMuriel S t mo de ppart
For that
the b o ok that Muriel likes S t mo de ind
a b o ok that Muriel to like S t mo de inf
the girl that reading the b o ok S t mo de ger
the b o ok that read by Muriel S t mo de ppart
Relative clause trees that have substitution of NP have the following feature equations
w
t hselect mo dei S t hmo dei NP
w r
NP t hselect mo dei ind
w
The examples that follow are intended to provide the rationale for the ab ove setting of
features
the b oy whose mother chased the cat S t hmo dei ind
r
the b oy whose mother to chase the cat S t hmo dei inf
r
the b oy whose mother eaten the cake S t hmo dei ppart
r
the b oy whose mother chasing the cat S t hmo dei ger
r
the b oy whose mother Bill b elieves to chase the cat
i i
S t hmo dei ind
r
The feature equations that app ear in trees whichhave substitution of PP are
w
S t hmo dei PP t hselect mo dei
r w
PP t hmo dei ind inf
w
Examples that justify the ab ove feature setting follow
the p erson by whom this machine was invented S t hmo dei ind
r
a baker in whom PRO to trust S t hmo dei inf
i i r
the fork with which Geo rey eaten the pudding S t h mo dei ppart
r
the p erson by whom this machine inventing S t hmo de i ger
r
As is the case for NP substitution any wh PP can substitute under PP This is implemented by the
w w
following equation
PP t hwhi
w
Not all cases of pied piping involve substitution of PP In some cases the P may b e built in In cases where
w
part of the pied pip ed PP is part of the anchor it continues to function as an anchor even after pied piping i e
the P no de and the NP no des are represented separately
w
CHAPTER RELATIVE CLAUSES
Further constraints on the null Comp
C
There are additional constraints on where the null Comp can o ccur The null Comp is not
C
p ermitted in cases of sub ject extraction unless there is an intervening clause or or the relative
clause is a reduced relative mo de ppart ger This can b e seen in
the toy likes Dafna
i C i
the toy Fred thinks likes Dafna
i C i
the b oy eating the guava
i C i
the guava eaten bytheboy
i C i
To mo del this paradigm the feature h rel proni is used in conjunction with the following
equations
S t hrel proni Comp t hrel proni
r
S b hrel proni S b hmo dei
r r
Comp b hrel proni ppart ger adj clause for
C
The full set of the equations shown ab ove is only present in Comp substitution trees involving
sub ject extraction So will not b e ruled out
the toy Dafna likes
i C i
The feature mismatch induced by the ab ove equations is not remedied by adjunction of just
any S adjunct b ecause all other S adjuncts are transparent to the hrel proni feature b ecause
of the following equation
S b hrel proni S t hrel proni
m
f
Reduced Relatives
Reduced relatives are p ermitted only in cases of sub ject extraction Past participial reduced
relatives are only p ermitted on passive clauses See
the toy playing the banjo
i C i
the instrument Amis playing
i C i
the day Amis playing the banjo
w C
the apple eaten by Dafna
i C i
the child the apple eaten by
i C i
the day Amis eaten the apple
w C
the apple Dafna eaten
i C i
the child eaten the apple
i C i
These restrictions are built into the mo de sp eci cations of S t So non passive cases of
sub ject extraction have the following feature equation
S t hmo dei ind ger inf
r
Passive cases of sub ject extraction have the following feature equation
S t hmo dei ind ger ppart inf
r
Finally all cases of non sub ject extraction have the following feature equation
S t hmo dei ind inf
r
Restrictive vs Non restrictive relatives
The English XTAG grammar do es not contain anysyntactic distinction b etween restrictiveand
non restrictive relatives b ecause we b elieve this to b e a semantic and or pragmatic di erence
External syntax
A relative clause can combine with the NP it mo di es in at least the following twoways
the toy Dafna likes
i C i
the toy Dafna likes
i C i
Based on cases like and whichare problematic for the structure in the
structure in is adopted
the man and the woman who met on the bus
the man and the woman who likeeach other
As it stands the RC analysis sketched so far will combine in twoways with the Determiner
tree shown in Figure giving us b oth the p ossiblities shown in and In order
to blo ck the structure exempli ed in the feature hrel clausei is used in combination with
the following equations
On the RC
NP b hrel clausei
r
On the Determiner tree
NP t hrel clausei
f
Together these equations blo ckintro duction of the determiner ab ove the relative clause
The determiner tree shown has the rel clause feature built in The RC analysis would givetwo parses
in the absence of this feature
CHAPTER RELATIVE CLAUSES
NPr
D NPf* displ-const : <1> NA conj : <2> case : <3> nom/acc agr : <4> rel-clause :- gerund : - wh : - quan : - gen : - definite : - decreas : - const : - card : -
the
Figure Determiner tree with rel clause feature Dnx
Other Issues
Interaction with adjoined Comps
The XTAG analysis now has two di erentways of intro ducing a complementizer like that or for
dep ending up on whether it o ccurs in a relativeclauseorinsentential complementation Relative
clause complementizers substitute in using the tree Comp while sentential complementizers
adjoin in using the tree COMPs Cases like where b oth kinds of complementizers
illicitly o ccur together are blo cked
that that Muriel wrote the b o ok
i w
i
This is accomplished by setting the S t comp feature in the relative clause tree to nil
r
tro duces the sentential complementa The S t comp feature of the auxiliary tree that in
r
tion that is set to that This leads to a feature clash ruling out On the other hand if a
sentential complement taking verbisadjoinedinatS this feature clash go es away cf
r
that Beth thinks that Muriel wrote the book
i w
i
Adjunction on PRO
Adjunction on PRO whichwould yield the ungrammatical is blo cked
Iwant PRO who Muriel likes to read a b o ok
This is done by sp ecifying the case feature of NP to b e nom acc The case feature
f
of PRO is null This leads to a feature clash and blo cks adjunction of relative clauses on to
PRO
Adjunct relative clauses
Twotyp es of trees to handle adjunct relative clauses exist in the XTAG grammar one in which
there is PP substitution with a null Comp built in and one in which there is a null NP
w w
built in and a Comp substitutes in There is no NP substitution tree with a null Comp
w
built in This is b ecause of the contrast between and
the day on whose predecessor Muriel left
C
the day whose predecessor Muriel left
C
In general adjunct relatives are not p ossible with an overt NP We do not consider
w
and to b e counterexamples to the ab ove statements b ecause we consider where and when
to b e exhaustive PPs that head a PP initial tree
the place where Muriel wrote her rst b o ok
C
the time when Muriel lived in Bryn Mawr
C
ECM
Cases where for assigns exceptional case cf are handled
a book for Muriel to read
w i
i
thetime for Muriel to leave Haverford
w
i
is implemented by a combination of the following equations The assignment of case by for
Comp t hassign casei acc
S t hassign casei Comp t hassign casei
r
S b hassign casei NP t hcasei
r
Cases not handled
Partial treatment of free relatives
Free relatives are only partially handled All free relatives on non sub ject p ositions and some
free relatives on sub ject p ositions are handled The structure assigned to free relatives treats the
extracted wh NP as the head NP of the relative clause The remaining relative clause mo di es
this extracted wh NP cf
Mary likes what ever
C i w
i
CHAPTER RELATIVE CLAUSES
where ever Mary lives
w C
who ever Muriel thinks likes Mary
w C i
i
However simple sub ject extractions without further emeb edding are not handled cf
who ever likes Bill
w C i
i
This is b ecause is treated exactly like the ungrammatical
the p erson likes Bill
w C i
i
Adjunct P stranding
The following cases of adjunct prep osition stranding are not handled cf
the p en Muriel wrote this letter with
the street Muriel lives on
Adjuncts are not built into elementary trees in XTAG So there is no clean way to represent
adjunct prep osition stranding A b etter solution is probably available if we make use of
multi comp onent adjunction
Overgeneration
The following ungrammatical sentences are currently b eing accepted by the XTAG grammar
This is b ecause no clean and conceptually attractive way of ruling them out is obvious to us
how as wh NP
In standard American English how is not acceptable as a relative pronoun cf
the way how PRO to solve this problem
C
However is accepted by the current grammar The only way to rule out would
be to intro duce a sp ecial feature devoted to this purp ose This is unapp ealing Further there
exist sp eech registers dialects of English where is acceptable
for trace e ects
is ungrammatical b eing an instance of a violation of the for trace lter of early transfor
mational grammar
the p erson for to read the b o ok
w i
i
The XTAG grammar currently accepts
It may b e of some interest that is acceptable in certain dialects of Belfast English
Internal head constraint
Relative clauses in English and in an overwhelming numb er of languages ob ey a no internal
head constraint This constraint is exempli ed in the contrast b etween and
the p erson who Muriel likes
i C i
the p erson which p erson Muriel likes
i C i
Weknowofnogoodway to rule out while still ruling in
the p erson whose mother Muriel likes
i C i
Dayal suggests that full NPs such as which person and whose mother are R
expressions while who and whose are pronouns R expressions unlike pronouns are sub ject
to Condition C is then ruled out as a violation of Condition C since the person and
which person are co indexed and the person c commands which person If we accept Dayal s ar
gument wehave a principled reason for allowing overgeneration of relative clauses that violate
the internal head constraint the reason b eing that the XTAG grammar do es generate binding
theory violations
Overt Comp constraint on stacked relatives
Stacked relatives of the kind in are handled
the b o ok that Bill likes which Mary wrote
There is a constraintonstacked relatives all but the relative clause closest to the head NP
must have either an overt Comp or an overt NP Thus is ungrammatical
w
the b o ok that Bill likes Mary wrote
Again no go o d way of handling this constraintisknown to us currently
Chapter
Adjunct Clauses
Adjunct clauses include sub ordinate clauses i e those with overt sub ordinating conjunctions
purp ose clauses and participial adjuncts
Sub ordinating conjunctions each select four trees allowing them to app ear in four di erent
p ositions relative to the matrix clause The p ositions are b efore the matrix clause after
the matrix clause b efore the VP surrounded bytwo punctuation marks and after the
matrix clause separated by a punctuation mark Each of these trees is shown in Figure
VP VP Sr r r Sr
VPf* PP Punct1↓ PP Punct2↓ VP* S1 Sf* NA NA Sf* Punct↓ PP NA NA
P1 S↓ P S↓
P S↓ P S↓ P N P1 P2
because in order as if when
Pss vxPNs puPPspuvx spuPs
Figure Auxiliary Trees for Sub ordinating Conjunctions
Sentence initial adjuncts adjoin at the ro ot S of the matrix clause while sentence nal
adjuncts adjoin at a VP no de In this the XTAG analysis follows the ndings on the attach
ment sites of adjunct clauses for conditional clauses Iatridou and for in nitival clauses
Browning One comp elling argument is based on Binding Condition C e ects As
can b e seen from examples below no Binding Condition violation o ccurs when the
adjunct is sentence initial but the sub ject of the matrix clause clearly governs the adjunct
clause when it is in sentence nal p osition and co indexation of the pronoun with the sub ject
of the adjunct clause is imp ossible
Unless she hurries Mary will b e late for the meeting
i i
She will b e late for the meeting unless Mary hurries
i i
Mary will b e late for the meeting unless she hurries
i i
We had previously treated sub ordinating conjunctions as a sub class of conjunction but
are now assigning them the POS preposition as there is such clear overlap between words
that function as prep ositions taking NP complements and sub ordinating conjunctions taking
clausal complements While there are some prep ositions whichonlytake NP complements and
some whichonlytake clausal complements manytake bothasshown in examples
and it seems to b e arti cial to assign them two di erent parts of sp eech
Helen left b efore the party
Helen left b efore the partybegan
Since the election Bill has b een elated
Since winning the election Bill has b een elated
Each sub ordinating conjunction selects the values of the mo de and comp features
of the sub ordinated S The mo de value constrains the typ es of clauses the sub ordinating
conjunction may app ear with and the comp value constrains the complementizers which
may adjoin to that clause For instance indicative sub ordinate clauses may app ear with the
complementizer that as in while participial clauses may not have any complementizers
Midge left that car so that Sam could drivetowork
Since that seeing the new VW Midge could think of nothing else
Multi word Sub ordinating Conjunctions
We extracted a list of multi word conjunctions suchas as if in order andfor al l that from
al For the most part the comp onents of the complex are all anchors as Quirk et
shown in Figures a In one case as ADVas there is a great deal of latitude in the choice
of adverb so this is a substitution site Figures b This multi anchor treatment is very
similar to that prop osed for idioms in Ab eill e and Schab es and the analysis of light
verbs in the XTAG grammar see section
Bare Adjunct Clauses
Bare adjunct clauses do not haveanovert sub ordinating conjunction but are typically parallel
in meaning to clauses with sub ordinating conjunctions For this reason we have elected to
handle them using the same trees shown ab ove but with null anchors They are selected at the
same time and in the same waythePRO tree is as they all have PRO sub jects Three values
of mo de are licensed inf in nitive ger gerundive and ppart past participal They
interact with complementizers as follows
We considered allowing bare indicative clauses suchas He died that others may live but these were considered
to o archaic to b e worth the additional ambiguitytheywould add to the grammar
CHAPTER ADJUNCT CLAUSES
VPr VPr
VPf* PP VPf* PP NA NA
P S↓ P S↓
P1 Ad P2 P1 Ad↓ P2
as soon as as as
a b
Figure Trees Anchored by Sub ordinating Conjunctions vxPARBPs and vxParbPs
Participial complements do not license any complementizers
Destroyed by the re the building still sto o d
The re raged for days destroying the building
That destroyed by the re the building still sto o d
In nitival adjuncts including purp ose clauses are licensed both with and without the
complementizer for
Harriet b ought a Mustang to impress Eugene
To impress Harriet Eugene dyed his hair
Tra c stopp ed for Harriet to cross the street
Discourse Conjunction
The CONJs auxiliary tree is used to handle discourse conjunction as in sentence Only
the co ordinating conjunctions and or and but are allowed to adjoin to the ro ots of matrix
sentences Discourse conjunction with and is shown in the derived tree in Figure
And Tru ula trees are what every one needs Seuss
While these sound a bit like extrap osed relative clauses see Kro ch and Joshi those move only to
the right and adjoin to S as these clauses are equally grammatical b oth sentence initially and sentence nally
we are analyzing them as adjunct clauses Sm Sr
Sr Sf NP VPm NA
DetP N VP Sr
NP VP NP VPr NA
PRO V PP DetP N Ad VP D fire VP PP NP VP1 NA NA
destroyed P NP D building still V the V P NP PRO V NP
by DetP N the stood raged for N destroying DetP N
D fire days D building
the the
a b
Figure Sample Participial Adjuncts
Sc
Conj Sr NA
and NP VPr
Nr V VP NA
N Nf are V NPr NA NA
Truffula trees ε NP Sr
N NP VP
what N V NP1 NA
everyone needs ε
Figure Example of discourse conjunction from Seuss The Lorax
Chapter
Imp eratives
Imp eratives in English do not require overt sub jects The sub ject in imp eratives is second
p erson i e you whether it is overt or not as is clear from the verbal agreement and the
interpretation Imp eratives with overt sub jects can be parsed using the trees already needed
for declaratives The imp erative cases in which the sub ject is not overt are handled by the
imp erative trees discussed in this section
The imp erative trees in English XTAG grammar are identical to the declarative tree except
that the NP sub ject p osition is lled byan the NP agr pers feature is set in the tree
to the value nd and the mo de feature on the ro ot no de has the value imp The value
for agr p ers is hardwired into the epsilon no de and insures the prop er verbal agreement
for an imp erative The mo de value of imp on the ro ot no de is recognized as a valid mo de
for a matrix clause The imp value for mo de also allows imp eratives to be blo cked from
app earing as emb edded clauses Figure is the imp erative tree for the transitive tree family S r displ-const : set1 : - assign-comp : inf_nil/ind_nil progressive : <11> perfect : <12> passive : <13> conditional : <14> assign-comp : <15> assign-case : <9> agr : <10> tense : <16> mode : <17> inv : - displ-const : set1 : <18> comp : nil extracted : -
NP 0 wh : - VP refl-obj : <8> NA case : <9> nom progressive : <11> agr : <10> num : plur/sing perfect : <12> 3rdsing : - passive : <13> pers : 2 conditional : <14> assign-comp : <15> assign-case : <9> agr : <10> tense : <16> pres mode : <17> displ-const : set1 : <18>
refl-obj : <1> mode : imp mainv : <2> tense : <3> assign-comp : <4> assign-case : <5> agr : <6> passive : <7> displ-const : set1 : -
ε ◊ ↓ V refl-obj : <1> NP 1 refl : <8> trans : + case : acc mainv : <2> tense : <3> assign-comp : <4> assign-case : <5> agr : <6> passive : <7> -
mode : base
Figure Transitive imp erativetree Inx Vnx
Chapter
Gerund NP s
There are two typ es of gerunds identi ed in the linguistics literature One is the class of
derived nominalizations also called nominal gerundives or action nominalizations exempli ed
in which instantiates the direct ob ject within an of PP The other is the class of so
called sentential or VP gerundives exempli ed in In the English XTAG grammar the
derived nominalizations are termed determiner gerunds and the sentential or VP gerunds
are termed NP gerunds
Some think that the selling of b onds is b ene cial
Are private markets approving of Washington bashing Wall Street
Both typ es of gerunds exhibit a similar distribution app earing in most places where NP s
are allowed The bold face p ortions of sentences show examples of gerunds as a
sub ject and as the ob ject of a prep osition
Avoiding such losses will take a monumental e ort
Mr Nolen s wandering do esn t make him a weirdo
Are private markets approving of Washington bashing Wall Street
The motivation for splitting the gerunds into two classes is semantic as well as structural
NP gerunds in nature Semantically the two gerunds are in sharp contrast with each other
refer to an action i e a way of doing something whereas determiner gerunds refer to a fact
Structurally there are a numb er of prop erties extensively discussed in Lees that show
that NP gerunds have the syntax of verbs whereas determiner gerunds have the syntax of basic
nouns Firstly the fact that the direct ob ject of the determiner gerund can only app ear within
an of PP suggests that the determiner gerund like nouns is not a case assigner and needs
insertion of the prep osition of for assignment of case to the direct ob ject NP gerunds like
verbs need no such insertion and can assign case to their direct ob ject Secondly like nouns
Thirdly only determiner gerunds can app ear with articles cf example and
determiner gerunds like nouns can be mo di ed by adjectives cf example whereas
an exception b eing the NP p ositions in equative BE sentences suchas John is my father
NP gerunds like verbs resist such mo di cation cf example Fourthly nouns unlike
verbs cannot co o ccur with asp ect cf example and Finally only NP gerunds
likeverbs can take adverbial mo di cation cf example and
the proving of the theorem det ger with article
the pro ving the theorem NP ger with article
John s rapid writing of the b o ok det ger with Adj
John s rapid writing the b o ok NP ger with Adj
John s ha ving written of the b o ok det ger with asp ect
John having written the b o ok NP ger with asp ect
His writing of the b o ok rapidly det ger with Adverb
His writing the b o ok rapidly NP ger with Adverb
In English XTAG the two typ es of gerunds are assigned separate trees within each tree
family but in order to capture their similar distributional b eha vior both are assigned NP as
the category lab el of their top no de The feature gerund distinguishes gerund NP s
from regular NP s where needed The determiner gerund and the NP gerund trees are discussed
in section and resp ectively
Determiner Gerunds
The determiner gerund tree in Figure is anchored by a V capturing the fact that the
gerund is derived from a verb The verb pro jects an N and instantiates the direct ob ject as an
of PP The nominal category pro jected by the verb can now display all the syntactic prop erties
of basic nouns as discussed ab ove For example it can be straightforwardly mo di ed by
adjectives it cannot co o ccur with asp ect and it can app ear with articles The only di erence
of the determiner gerund nominal with the basic nominals is that the former cannot o ccur
without the determiner whereas the latter can The determiner gerund tree therefore has an
initial D mo difying the N It is used for gerunds such as the ones in b old face in sentences
and
The D no de can take a simple determiner cf example a genitive pronoun cf
example or a genitiveNP cf example
This feature is also needed to restrict the selection of gerunds in NP p ositions For example the sub ject
and ob ject NP s in the equative BE tree Tnx BEnx cannot b e lled by gerunds and are therefore assigned
the feature gerund which prevents gerunds whichhave the feature gerund from substituting into
these NP p ositions
Note that the determiner can adjoin to the gerund only from within the gerund tree Adjunction of deter
miners to the gerund ro ot no de is prevented by constraining determiners to only select NP s with the feature
gerund This rules out sentences like Private markets approved of the the sel ling of bonds
The trees for genitive pronouns and genitiveNP shave the ro ot no de lab elled as D b ecause they seem to
function as determiners and as such sequence with the rest of the determiners See Chapter for discussion on
determiner trees
CHAPTER GERUND NP S
NP agr : 3rdsing : + pers : 3 num : sing case : nom/acc wh : <3> decreas : <4> gen : <5> card : <6> quan : <7> definite : <8> const : <9>
D↓ wh : <3> N decreas : <4> gen : <5> card : <6> quan : <7> definite : <8> const : <9>
V◊ mode : ger PP1 wh : <1> assign-case : <2>
P1 assign-case : <2> NP1↓ wh : <1> assign-case : acc case : <2>
of
Dnx Vnx
Figure Determiner Gerund tree from the transitive tree family Dnx Vnx
Some think that the selling of b onds is b ene cial
His painting of Mona Lisa is highly acclaimed
Are private markets approving of Washington s bashing of Wall Street
NP Gerunds
NP gerunds showanumb er of structural p eculiarities the crucial one b eing that they have the
internal prop erties of sentences In the English XTAG grammar we adopt a p osition similar
to that of Rosenbaum and Emonds that these gerunds are NP s exhaustively
dominating a clause Consequently the tree assigned to the transitive NP gerund tree cf
Figure lo oks exactly like the declarative transitive tree for clauses except for the ro ot
no de lab el and the feature values The anchoring verb pro jects a VP Auxiliary adjunction is
allowed sub ject to one constraint that the highest verbintheverbal sequence b e in gerundive
form regardless of whether it is a main or auxiliary verb This constraint is implemented by
requiring the topmost VP no de to b e mo de ger In the absence of any adjunction the
anchoring verb itself is forced to be gerundive But if the verbal sequence has more than one
verb then the sequence and form of the verbs is limited by the restrictions that eachverb in the
sequence imp oses on the succeeding verb The nature of these restrictions for sentential clauses
and the manner in which they are implemented in XTAG are both discussed in Chapter
The analysis and implementation discussed there di ers from that required for gerunds only in
one resp ect that the highest verb in the verbal sequence is required to b e mo de ger
NPr gerund : + displ-const : set1 : <3> agr : 3rdsing : + pers : 3 num : sing case : nom/acc wh : <4>
NP0↓ wh : <4> VP mode : ger case : acc/none/gen displ-const : set1 : <3> - passive : <1> mode : <2> compar : - displ-const : set1 : -
V◊ passive : <1> - NP1↓ case : acc
mode : <2>
Gnx Vnx
Figure NP Gerund tree from the transitive tree family Gnx Vnx
Additionally the sub ject in the NP gerund tree is required to have case acc none gen
i e it can b e either a PRO cf example a genitiveNP cf example or an accusative
NP cf example The whole NP formed by the gerund can o ccur in either nominativeor
accusative p ositions
CHAPTER GERUND NP S
John do es not like wearing a hat
Are private markets approving of Washington s bashing Wall Street
Mother disapproved of me wearing such casual clothes
One question that arises with resp ect to gerunds is whether there is anything sp ecial ab out
their distribution as compared to other typ es of NP s In fact it app ears that gerund NP s
can o ccur in anyNP p osition Some verbs might not seem to b e very accepting of gerund NP
arguments as in b elow but we b elieve this to b e a semantic incompatibility rather than
asyntactic problem since the same structures are p ossible with other lexical items
John s tinkering ran
NP NP
John s tinkering worked
NP NP
By having the ro ot no de of gerund trees b e NP the gerunds have the same distribution as
any other NP in the English XTAG grammar without doing anything exceptional The clause
structure is captured b y the form of the trees and by inclusion in the tree families
Gerund Passives
It was mentioned ab ove that the NP gerunds display certain clausal prop erties It is therefore
not surprising that they to o have their own set of transformationally related structures For
example NP gerunds allow passivization just like their sentential counterparts cf examples
and
The lawyers ob jected to the slanderous book being written by John
Susan could not forget having been embarrassed by the vicar
In the English XTAG grammar gerund passives are treated in an almost exactly similar
fashion to sentential passives and are assigned separate trees within the appropriate tree fam
ilies The passives o ccur in pairs one with the by phrase and another without it There are
two feature restrictions imp osed on the passive trees a only verbs with mo de ppart
i e verbs with passive morphology can b e the anchors and b the highest verb in the verb
sequence is required to b e mo de ger The two restrictions together ensure the selection
of only those sequences of auxiliary verb s that select mo de ppart and passive
suchas being or having been but NOT having The passive trees are assumed to b e related
to only the NP gerund trees and not the determiner gerund trees since passive structures
involve movement of some ob ject to the sub ject p osition in a movement analysis Also like
the sentential passives gerund passives are found in most tree families that have a direct ob ject
in the declarativetree Figure shows the gerund passive trees for the transitive tree family NPr NPr gerund : + wh : <6> gerund : + displ-const : set1 : <5> wh : <4> agr : 3rdsing : + displ-const : set1 : <3> pers : 3 agr : 3rdsing : + num : sing case : nom/acc pers : 3 num : sing case : nom/acc
NP1↓ wh : <6> VP mode : ger case : acc/none/gen displ-const : set1 : <5> - compar : - displ-const : set1 : - passive : <3> NP1↓ case : acc/none/gen VP mode : ger mode : <4> wh : <4> displ-const : set1 : <3> - passive : <1> mode : <2> V◊ trans : + PP0 passive : <3> + wh : <1> displ-const : set1 : - mode : <4> ppart assign-case : <2> compar : -
P0 assign-case : <2> NP0↓ wh : <1> V◊ assign-case : acc case : <2> trans : + passive : <1> + mode : <2> ppart
by
a Gnx Vbynx b Gnx V
Figure Passive Gerund trees from the transitive tree family Gnx Vbynx a and
Gnx V b
CHAPTER GERUND NP S
Part IV
Other Constructions
Chapter
Determiners and Noun Phrases
In our English XTAG grammar all nouns select the noun phrase NP tree structure shown
in Figure Common nouns do not require determiners in order to form grammatical
NPs Rather than b eing ungrammatical singular countable nouns without determiners are
restricted in interpretation and can only b e interpreted as mass nouns Allowing all nouns to
head determinerless NPs correctly treats the individuation in countable NPs as a prop erty of
determiners Common nouns havenegative values for determiner features in the lexicon in
our analysis and can only acquire a p ositive value for those features if determiners adjoin
to them Other typ es of NPs such as pronouns and prop er nouns have b een argued by Abney
e to the determiner p osition b ecause they Abney to either be determiners or to mov
exhibit determiner like b ehavior We can capture this insight in our system by giving pronouns
and prop er nouns p ositive values for determiner features For example pronouns and prop er
nouns would b e marked as de nite a value that NPs containing common nouns can only obtain
byhaving a de nite determiner adjoin In addition to the determiner features nouns also have
values for features such as re exive re case pronoun pron and conjunction conj
A single tree structure is selected by simple determiners an auxiliary tree which adjoins
to NP An example of this determiner tree anchored by the determiner these is shown in
Figure In addition to the determiner features the tree in Figure has noun features
such as case see section the conj feature to control conjunction see Chapter
rel clause see Chapter and gerund see Chapter which prevent determiners from
adjoining on top of relative clauses and gerund NPs resp ectively and the displ const feature
which is used to simulate multi comp onent adjunction
Complex determiners such as genitives and partitives also anchor tree structures that adjoin
to NP They di er from the simple determiners in their internal complexity Details of our
treatment of these more complex constructions app ear in Sections and Sequences of
determiners as in the NPs al l her dogs or those ve dogs are derived bymultiple adjunctions
of the determiner tree with each tree anchored by one of the determiners in the sequence The
order in which the determiner trees can adjoin is controlled by features
adjoining onto NPs is similar to that of Ab eill e This treatment of determiners as
and allows us to capture one of the insights of the DP hyp othesis namely that determiners
select NPs as complements In Figure the determiner and its NP complement app ear in
A more detailed discussion of this analysis can b e found in Ho ckey and Mateyak
CHAPTER DETERMINERS AND NOUN PHRASES
NP compl : <1> gen : <2> definite : <3> decreas : <4> quan : <5> const : <6> card : <7> conj : <8> pron : <9> wh : <10> case : <11> refl : <12> agr : <13>
N◊ compl : <1> gen : <2> definite : <3> decreas : <4> quan : <5> const : <6> card : <7> conj : <8> pron : <9> wh : <10> case : <11> nom/acc refl : <12>
agr : <13>
Figure NP Tree
the con guration that is typically used in LTAG to represent selectional relationships That
is the head serves as the anchor of the tree and it s complement is a sister no de in the same
elementary tree
The XTAG treatment of determiners uses nine features for representing their prop erties
de niteness de nite quantity quan cardinality card genitive gen decreasing de
compl Seven of these creas constancy const wh agreement agr and complement
features were develop ed by semanticists for their accounts of semantic phenomena Keenan
and Stavi Barwise and Co op er Partee et al another was develop ed for NP r wh : <1> decreas : <2> compl : <3> gen : <4> card : <5> quan : <6> definite : <7> const : <8> agr : <9> case : <10> nom/acc conj : <11> displ-const : <12>
D wh : <1> NP f* gerund : - decreas : <2> NA rel-clause : - compl : <3> agr : <9> gen : <4> case : <10> card : <5> conj : <11> quan : <6> displ-const : <12> definite : <7> const : <8> wh : -
These
Figure Determiner Trees with Features
a semantic account of determiner negation by one of the authors of this determiner analysis
Mateyak and the last is the familiar agreement feature When used together these
features also account for a substantial p ortion of the complex patterns of English determiner
sequencing Although we do not claim to have exhaustively covered the sequencing of deter
miners in English we do cover a large subset b oth in terms of the phenomena handled and
in terms of corpus coverage The XTAG grammar has also been extended to include complex
determiner constructions such as genitives and partitives using these determiner features
Each determiner carries with it a set of values for these features that represents its own
prop erties and a set of values for the prop erties of NPs to which can adjoin The features are
crucial to ordering determiners correctly The semantic de nitions underlying the features are
given b elow
De niteness Possible Values
CHAPTER DETERMINERS AND NOUN PHRASES
A function f is de nite i f is non trivial and whenever f s then it is always the
intersection of one or more individuals Keenan and Stavi
Quantity Possible Values
If A and B are sets denoting an NP and asso ciated predicate resp ectively E is a domain
in a mo del M and F is a bijection from M to M thenwesay that a determiner satis es
the constraint of quantityifDet AB Det F A F B Partee et al
E E
Cardinality Possible Values
A determiner D is cardinal i D cardinal numbers
Genitive Possible Values
Possessive pronouns and the p ossessive morpheme s are marked gen all other nouns
are gen
Decreasing Possible Values
A set of Q prop erties is decreasing i whenever s t and t Q then s Q A function f is
decreasing i for all prop erties f s is a decreasing set
A non trivial NP one with a Det is decreasing i its denotation in any mo del is decreas
ing Keenan and Stavi
Constancy Possible Values
If A and B are sets denoting an NP and asso ciated predicate resp ectively and E is
a domain then we say that a determiner displays constancy if A B E E then
AB Mo di ed from Partee et al Det AB Det
E
E
Complement Possible Values
there exists a A determiner Q is p ositive complement if and only if for every set X
continuous set of p ossible values for the size of the negated determined set NOT QX
and the cardinalityofQXistheonly asp ect of QX that can b e negated adapted from
Mateyak
The wh feature has been discussed in the linguistics literature mainly in relation to wh
movement and with resp ect to NPs and nouns as well as determiners We give a shallow but
useful working de nition of the wh feature b elow
Wh Possible Values
Interrogative determiners are wh all other determiners are wh
The agr feature is inherently a noun feature While determiners are not morphologically
marked for agreement in English many of them are sensitive to number Many determiners
are semantically either singular or plural and must adjoin to nouns which are the same For
example a canonlyadjointosingularnouns a dog vs a dogs while many must have plurals
many dogs vs many dog Other determiners such as some are unsp eci ed for agreement in
our analysis b ecause they are compatible with either singulars or plurals some dog some dogs
The p ossible values of agreement for determiners are sg pl
Det de nite quan card gen wh decreas const agr compl
all pl
b oth pl
this sg
these pl
that sg
those pl
what
whatever
which
whichever
the
each sg
every sg
a an sg
some
some pl
any sg
another sg
few pl
afew pl
many pl
many a an sg
several pl
various pl
sundry pl
no
neither
either
GENITIVE UN
CARDINAL pl
PARTITIVE UN
Table Determiner Features asso ciated with D anchors
CHAPTER DETERMINERS AND NOUN PHRASES
The determiner tree in Figure shows the appropriate feature values for the determiner
these whileTable shows the corresp onding feature values of several other common deter
miners
In addition to the features that represent their own prop erties determiners also have features
to represent the selectional restrictions they imp ose on the NPs they take as complements The
selectional restriction features of a determiner app ear on the NP fo otno de of the auxiliary
tree that the determiner anchors The NP no de in Figure shows the selectional feature
f
restriction imp osed by these while Table shows the corresp onding selectional feature
restrictions of several other determiners
The Wh Feature
A determiner with a wh feature is always the left most determiner in linear order since
no determiners have selectional restrictions that allow them to adjoin onto an NP with a wh
feature value The presence of a wh determiner mak es the entire NP wh and this is correctly
represented by the coindexation of the determiner and ro ot NP no des values for the wh feature
Wh determiners selectional restrictions on the NP fo ot no de of their tree only allows them
adjoin onto NPs that are wh or unsp eci ed for the wh feature Therefore ungrammatical
sequences suchas which what dog are imp ossible The adjunction of wh determiners onto
wh pronouns is also prevented by the same mechanism
Multi word Determiners
The system recognizes the multi word determiners afewand many a The features for a multi
word determiner are lo cated on the parent no de of its two comp onents see Figure We
chose to represent these determiners as multi word constituents b ecause neither determiner
retains the same set of features as either of its parts For example the determiner a is sg and
few is decreasing while a few is pl and increasing Additionally many is pl and a displays
constancy but many a is sg and do es not display constancy Example sentences app ear in
Multi word Determiners
a few teasp o ons of sugar should b e adequate
many a man has attempted that stunt but none have succeeded
We use the symb ol UN to represent the fact that the selectional restrictions for a given feature are unsp eci ed
meaning the noun phrase that the determiner selects can b e either p ositive or negative for this feature
Except one whichis sg
Except one whichis compl
A partitive can b e either quan or quan dep ending up on the nature of the noun that anchors the partitive
If the anchor noun is mo di ed then the quantity feature is determined by the mo di er s quantityvalue
In addition to this tree these would also anchor another auxiliary tree that adjoins onto card determiners
one di ers from the rest of CARD in selecting singular nouns
Det de n quan card gen wh decreas const agr compl e g
pl dogs
all UN UN UN pl these dogs
UN UN UN UN UN UN pl UN ve dogs
pl dogs
both UN UN UN pl these dogs
sg dog
UN UN UN few dogs
this that UN UN pl many dogs
UN UN UN UN UN UN sg UN ve dogs
pl dogs
these those UN UN pl UN few dogs
UN UN UN UN UN UN pl UN ve dogs
what which dog s
whatever UN UN UN few dogs
whichever UN UN UN UN UN UN UN many dogs
dog s
the UN UN UN few dogs
UN the me
UN UN pl many dogs
UN UN UN UN UN UN UN ve dogs
sg dog
every each UN UN UN few dogs
UN UN UN UN UN UN UN ve dogs
a an sg dog
some dog s
some UN UN UN UN UN UN pl UN dogs
sg dog
any UN UN UN few dogs
UN UN UN UN UN UN UN ve dogs
sg dog
another UN UN UN few dogs
UN UN UN UN UN UN UN ve dogs
few pl dogs
afew pl dogs
many pl dogs
many a an sg dog
several pl dogs
various pl dogs
sundry pl dogs
no dog s
neither sg dog
either sg dog
Table Selectional Restrictions Imp osed by Determiners on the NP fo ot no de
CHAPTER DETERMINERS AND NOUN PHRASES
Det de nite quan card gen wh decreas const agr compl
UN UN UN
GENITIVE UN UN pl
UN UN UN UN UN UN UN
pl
pl
CARDINAL pl
PARTITIVE UN UN UN UN UN UN UN UN
Table Selectional Restrictions Imp osed by Groups of Determiners Determiner Construc tions
NPr
D const : <5> NPf * displ-const : <1> definite : <6> NA conj : <2> quan : <7> case : <3> nom/acc card : <8> agr : <4> gen : <9> wh : - decreas : <10> quan : - wh : <11> gen : - agr : <4> definite : - wh : - decreas : - quan : + const : - predet : - card : - gen : - definite : - decreas : - card : - agr : 3rdsing : + pers : 3 num: sing
D1 D2
many a
Figure Multi word Determiner tree DDnx
Genitive Constructions
There are two kinds of genitive constructions genitive pronouns and genitive NP s which
have an explicit genitive marker s asso ciated with them It is clear from examples such as
her dog returned home and her ve dogs returned home vs dog returned home that genitive
pronouns function as determiners and as such they sequence with the rest of the determiners
The features for the genitives are the same as for other determiners Genitives are not required
to agree with either the determiners or the nouns in the NPs that they mo dify The value of
the agr feature for an NP with a genitive determiner dep ends on the NP to which the genitive
determiner adjoins While it might seem to make sense to take their as pl my as sg and
Alfonso s as sg this number and person information only e ects the genitive NP itself and
b ears no relationship to the number and p erson of the NPs with these items as determiners
Consequently wehave represented agr as unsp eci ed for genitives in Table
Genitive NP s are particularly interesting b ecause they are p otentially recursive structures
Complex NP s can easily b e emb edded within a determiner
John s friend from high scho ol s uncle s mother came to town
There are two things to note in the ab ove example One is that in emb edded NPs the
genitive morpheme comes at the end of the NP phrase even if the head of the NP is at the
b eginning of the phrase The other is that the determiner of an emb edded NP can also be a
genitiveNP hence the p ossibility of recursive structures
In the XTAG grammar the genitive marker s is separated from the lexical item that it
is attached to and given its own category G In this way we can allow the full complexity
of NP s to come from the existing NP system including any recursive structures As with the
simple determiners there is one auxiliary tree structure for genitives which adjoins to NPs As
can b e seen in this tree is anchored by the genitivemarker s and has a branching D no de
which accomo dates the additional internal structure of genitive determiners Also like simple
determiners there is one initial tree structure Figure available for substitution where
needed as in for example the Determiner Gerund NP tree see Chapter for discussion on
determiners for gerund NP s
Since the NP no de which is sister to the G no de could also have a genitive determiner in it
the typ e of genitive recursion shown in is quite naturally accounted for by the genitive
tree structure used in our analysis
Partitive Constructions
The deciding factor for including an analysis of partitive constructions e g some kind of al l
of as a complex determiner constructions was the behavior of the agreement features If
partitive constructions are analyzed as an NP with an adjoined PP then we would exp ect to
get agreement with the head of the NP as in If on the other hand we analyze them
as a determiner construction then we would exp ect to get agreement with the noun that the
determiner sequence mo di es as we do in
a kind of these machines is prone to failure
a kind of these machines are prone to failure
In other words for partitive constructions the semantic head of the NP is the second rather
than the rst noun in linear order That the agreement shown in is p ossible suggests that
CHAPTER DETERMINERS AND NOUN PHRASES
NP r decreas : <2> gen : <3> card : <4> quan : <5> definite : <6> const : <7> agr : <1> wh : <8>
D decreas : <2> NP f* agr : <1> gen : <3> card : <4> quan : <5> definite : <6> const : <7> wh : <8> gen : <9> wh : <10>
NP↓ wh : <10> G◊ gen : <9>
case : nom/acc gen : +
Figure Genitive Determiner Tree
D gen : <1> wh : <2>
NP↓ wh : <2> G◊ gen : <1>
case : nom/acc gen : +
Figure Genitive NP tree for substitution DnxG
the second noun in linear order in these constructions should also be treated as the syntactic
head Note that b oth the partitive and PP readings are usually p ossible for a particular NP In
the cases where either the partitive or the PP reading is preferred we take it to b e just that
a preference most appropriately mo deled not in the grammar but in a comp onentsuch as the
heuristics used with the XTAG parser for reducing the analyses pro duced to the most likely
In our analysis the partitive tree in Figure is anchored by one of a limited group of
nouns that can app ear in the determiner p ortion of a partitive construction A rough semantic
characterization of these nouns is that they either representquantity e g part half most pot
cup pound etc or classi cation e g type variety kind version etc In the absence of a
more implementable characterization we use a list of such nouns compiled from a descriptive
grammar Quirk et al a thesaurus and from online corp ora In our grammar the nouns
on the list are the only ones that select the partitive determiner tree
Like other determiners partitiv es can adjoin to an NP consisting of just a noun a certain
kind of machine or adjoin to NPs that already have determiners some parts of these
machines Notice that just as for the genitives the complexity and the recursion are contained
b elow the D no de and rest of the structure is the same as for simple determiners
Adverbs Noun Phrases and Determiners
Many adverbs interact with the noun phrase and determiner system in English For example
consider sentences below
Approximately thirty p eople came to the lecture
Practically every p erson in the theater was laughing hysterically during that scene
Only John s crazy mother can make stu ng that tastes so go o d
Relatively few programmers remember how to program in COBOL
Not every martian would p ostulate that all humans sp eak a universal language
Enough money was gathered to pay o the group gift
Quite a few burglaries o ccurred in that neighb orho o d last year
Iwanted to b e paid double the amount they o ered
Although there is some debate in the literature as to whether these should b e classi ed as
determiners or adverbs we b elieve that these items that interact with the NP and determiner
system are in fact adverbs These items exhibit a broader distribution than either determiners
or adjectives in that they can mo dify many other phrasal categories including adjectives verb
phrases prep ositional phrases and other adverbs
Using the determiner feature system we can obtain a close approximation to an accurate
characterization of the b ehavior of the adverbs that interact with noun phrases and determiners
Adverbs can adjoin to either a determiner or a noun phrase see gure with the adverbs
restricting what typ es of NPs or determiners they can mo dify by imp osing feature requirements
on the fo ot D or NP no de For example the adverb approximately seen in ab ove selects
CHAPTER DETERMINERS AND NOUN PHRASES
NP r NA decreas : <5> compl : <6> gen : <7> card : <8> quan : <9> definite : <10> const : <11> agr : <13> wh : <12> case : <14> conj : <15> displ-const : <16>
D decreas : <5> NP f* rel-clause : - compl : <6> NA agr : <13> gen : <7> case : <14> nom/acc card : <8> conj : <15> quan : <9> displ-const : <16> definite : <10> const : <11> wh : <12> wh : <4>
NP wh : <4> P wh : <1> case : <2> agr : <3>
N◊ wh : <1> of case : <2>
agr : <3>
Figure Partitive Determiner Tree
for determiners that are card The adverb enough in is an example of an adverb that
selects for a noun phrase sp eci cally a noun phrase that is not mo di ed by a determiner
Most of the adverbs that mo dify determiners and NPs divide into six classes with some
minor variation within classes based on the pattern of these restrictions Three of the classes
are adverbs that mo dify determiners while the other three mo dify NPs NPr NPr
Ad NPf Dr NPf NA double D NPf NA Ad Df N NA the N
approximately thirty people amount
a b
Figure a Adverb mo difying a determiner b Adverb mo difying a noun phrase
The largest of the ve classes is the class of adverbs that mo dify cardinal determiners This
class includes among others the adverbs about at most exactly nearly and only These ad
verbs have the single restriction that they must adjoin to determiners that are card Another
class of adverbs consists of those that can mo dify the determiners every al l any and no The
adverbs in this class are almost nearly and practical ly Closely related to this class are the ad
verbs mostly and roughly which are restricted to mo difying every and al l andhardly whichcan
only mo dify any To select for every al l andany these adverbs select for determiners that are
quan card const compl and to select for no the adverbs cho ose a determiner that
is quan decreas const The third class of adverbs that mo dify determiners are those
that mo dify the determiners few and many representable by the feature sequences quan
decreas const and quan decreas const pl compl resp ectively Examples of
these adverbs are awful ly fairly relatively andvery
Of the three classes of adverbs that mo dify noun phrases one actually consists of a single
adverb not that only mo di es determiners that are compl Another class consists of the
fo cus adverbs at least even only andjust These adverbs select NPs that are wh and card
For the NPs that are card the fo cus adverbs actually mo dify the cardinal determiner and so
these adverbs are also included in the rst class of adverbs mentioned in the previous paragraph
The last ma jor class that mo dify NPs consist of the adverbs double and twice which select NPs
that are de nite i e the this that those these and the genitives
Although these restrictions succeed in recognizing the correct determiner adverb sequences
a few unacceptable sequences slip through For example in handling the second class of adverbs
mentioned ab ove every al l and any share the features quan card const compl
with a and another and so nearly a man is acceptable in this system In addition to this
over generation within a ma jor class the adverb quite selects for determiners and NPs in what
seems to b e a purely idiosyncratic fashion Consider the following examples
a Quite a few memb ers of the audience had to leave
y new participants at this year s conference b There were quite man
c Quite few triple jump ers have jump ed that far
CHAPTER DETERMINERS AND NOUN PHRASES
d Taking the dayo was quite the right thing to do
e The recent negotiation asco is quite another issue
f Pandora is quite a cat
In examples a c quite mo di es the determiner while in d f quite mo d
i es the entire noun phrase Clearly it functions in a di erent manner in the two sets of
sentences in a c quite intensi es the amount implied by the determiner whereas in
d f it singles out an individual from the larger set to which it b elongs To capture the
selectional restrictions needed for a c we utilize the two sets of features mentioned
previously for selecting few and many However afewcannot b e singled out so easily using the
sequence quan card decreas const pl compl we also accept the ungrammatical
NPs quite several members and quite some members where quite mo di es some In selecting
the as in d with the features de nite gen sg quite also selects this and that which
are ungrammatical in this p osition Examples e and f present yet another obstacle
in that in selecting another and a quite erroneously selects every and any
It may b e that there is an undiscovered semantic feature that would alleviate these di cul
ties However on the whole the determiner feature system wehave prop osed can b e used as a
surprisingly e cient metho d of characterizing the interaction of adverbs with determiners and
noun phrases
Chapter
Mo di ers
This chapter covers various typ es of mo di ers adverbs prep ositions adjectives and noun
mo di ers in noun noun comp ounds These categories optionally mo dify other lexical items
and phrases by adjoining onto them In their mo di er function these items are adjuncts they
are not part of the sub categorization frame of the items they mo dify Examples of some of
these mo di ers are shown in
certainly the Octob er sell o didn t settle any stomachs WSJ
AD V AD V
Mr Bakes previously had a turn at running Continental WSJ
AD V AD V
most foreign government bond prices rose during the w eek
AD J AD J N N N N PP
PP
The trees used for the various mo di ers are quite similar in form The mo di er anchors
the tree and the ro ot and fo ot no des of the tree are of the category that the particular anchor
mo di es Some mo di ers e g prep ositions select for their own arguments and those are also
included in the tree The fo ot no de may be to the right or the left of the anchoring mo di er
and its arguments dep ending on whether that mo di er o ccurs b efore or after the category it
mo di es For example almost all adjectives app ear to the left of the nouns they mo dify while
prep ositions app ear to the right when mo difying nouns
Adjectives
In addition to b eing mo di ers adjectives in the XTAG English grammar can be also anchor
clauses see Adjective Small Clauses in Chapter There is also one tree family Intransitive
that has an adjective as an argument and is used for sentences with Adjective Tnx Vax
such as Seth felt happy In that tree family the adjective substitutes into the tree rather than
adjoining as is the case for mo di ers
As mo di ers adjectives anchor the tree shown in Figure The features of the N no de
onto which the An tree adjoins are passed through to the top no de of the resulting N The
null adjunction marker NA on the N fo ot no de imp oses right binary branching such that each
Relative clauses are discussed in Chapter
CHAPTER MODIFIERS
Nr gen : <1> definite : <2> decreas : <3> quan : <4> const : <5> card : <6> conj : <7> displ-const : <8> wh : <9> pron : <10> assign-comp : <11> agr : <12> case : <13>
A◊ Nf* gen : <1> NA definite : <2> decreas : <3> quan : <4> const : <5> card : <6> conj : <7> displ-const : <8> wh : <9> pron : <10> assign-comp : <11> agr : <12>
case : <13>
Figure Standard Tree for Adjective mo difying a Noun An
subsequent adjective must adjoin on top of the leftmost adjective that has already adjoined
Due to the NA constraint a sequence of adjectives will have only one derivationintheXTAG
grammar The adjective s morphological features such as sup erlative or comparative are in
stantiated by the morphological analyzer See Chapter for a description of how we handle
comparatives At this point the treatment of adjectives in the XTAG English grammar do es
onto not include selectional or ordering restrictions Consequently any adjective can adjoin
any noun and on top of any other adjective already mo difying a noun All of the mo di ed noun
phrases shown in currently parse with the same structure shown for colorless green
ideas in Figure
big green bugs
big green ideas
colorless green ideas
green big ideas
NP
Nr
A Nf NA
colorless A Nf NA
green ideas
Figure Multiple adjectives mo difying a noun
While are all semantically anomalous also su ers from an ordering prob
lem that makes it seem ungrammatical in the absence of any licensing context One could
argue that the grammar should accept but not One of the future goals for
the grammar is to develop a treatment of adjective ordering similar to that develop ed by Hockey
and Mateyak for determiners An adequate implementation of ordering restrictions for
adjectives would rule out
Noun Noun Mo di ers
Noun noun comp ounding in the English XTAG grammar is very similar to adjective noun
mo di cation The noun mo di er tree shown in Figure has essentially the same structure
as the adjective mo di er tree in Figure except for the syntactic category lab el of the
anchor
See Chapter or Ho ckey and Mateyak for details of the determiner analysis
CHAPTER MODIFIERS
Nr assign-comp : <1> displ-const : <2> pron : <3> - wh : <4> agr : <5> case : <6> nom/acc
N◊ case : nom/acc Nf* assign-comp : <1> pron : - NA displ-const : <2> wh : <4> agr : <5> case : <6>
pron : <3>
Figure Noun noun comp ounding tree Nn not all features displayed
Noun comp ounds have a variety of scop e p ossibilities not available to adjectives as illus
trated by the single bracketing p ossibility in and the two p ossibilities for This
ambiguity is manifested in the XTAG grammar by the two p ossible adjunction sites in the
noun noun comp ound tree itself Subsequent mo difying nouns can adjoin either onto the N
r
keting p os no de or onto the Nanchor no de of that tree which results in exactly the two brac
sibilities shown in This inherent structural ambiguityresultsin noun noun comp ounds
regularly having multiple derivations However the multiple derivations are not a defect in
the grammar b ecause they are necessary to correctly represent the genuine ambiguity of these
phrases
big green design
N N N N
computer furniture design
N N N N
computer furniture design
N N N N
Noun noun comp ounds have no restriction on numb er XTAG allows nouns to be either
singular or plural as in
Hyun is taking an algorithms course
wa es are in the frozen fo o ds section
I enjoy the dog shows
Time Noun Phrases
Although in general NPs cannot mo dify clauses or other NPs there is a class of NPs with
meanings that relate to time that can do so We call this class of NPs Time NPs Time NPs
b ehave essentially like PPs Like PPs time NPs can adjoin at four places to the right of an
NP totheright and left of a VP and to the left of an S
Time NPs may include determiners as in this month in example or may be single
lexical items as in today in example Like other NPs time NPs can also include adjectives
as in example
Elvis left the building this week
Elvis left the building to day
It has no b earing on our work force to day WSJ
claimed twolives The re yesterday
Today it has no b earing on our work force
Michael late yesterday announced a buy back program
The XTAG analysis for time NPs is fairly simple requiring only the creation of prop er NP
auxiliary trees Only nouns that can b e part of time NPs will select the relevant auxiliary trees
and so only these typ e of NPs will b ehavelike PPs under the XTAG analysis Currently about
words select Time NP trees but since these words can form NPs that include determiners
and adjectives a large numb er of phrases are covered by this class of mo di ers
Corresp onding to the four p ositions listed ab ove time NPs can select one of the four trees
shown in Figure
Determiners can b e added to time NPs by adjunction in the same way that they are added
to NPs in other p ositions The trees in Figure show that the structures of examples
and di er only in the adjunction of this to the time NP in example
The sentence
Esso said the Whiting eld started pro duction Tuesday WSJ
has at least two di erentinterpretations dep ending on whether Tuesday attaches to said or
to started Valid time NP analyses are available for b oth these interpretations and are shown
in Figure
Derived tree structures for examples whichshow the four p ossible time NP
p ositions are shown in Figures and The derivation tree for example is also
shown in Figure
There maybe other classes of NPs such as directional phrases such as north south etc which behave
similarly Wehavenotyet analyzed these phrases
CHAPTER MODIFIERS
Sr VPr VPr NPr
NP S* NP VP* VP* NP NPf * NP NA NA NA NA
N◊ N◊ N◊ N◊
Ns Nvx vxN nxN
Figure Time Phrase Mo di er trees Ns Nvx vxN nxN
S r Sr
NP VP r NP VPr
N VP NP r N VP NP NA NA
Elvis V NP r D NP f Elvis V NPr N NA
left D NPf today left D NP f this N NA NA
the N week the N
building building
Figure Time NPs with and without a determiner
Prep ositions
There are three basic typ es of prep ositional phrases and three places at which they can adjoin
The three typ es of prep ositional phrases are Prep osition with NP Complement Prep osition
with Sentential Complement and Exhaustive Prep osition The three places are to the rightof
an NP to the rightofaVP andtotheleftofanS Each of the three typ es of PP can adjoin at
each of these three places for a total of nine PP mo di er trees Table gives the tree family
names for the various combinations of typ e and lo cation See Section for discussion of
the spuPnx which handles p ost sentential comma separated PPs
The subset of prep osition anchored mo di er trees in Figure illustrates the lo cations and
the four PP typ es Example sentences using the trees in Figure are shown in
There are also more trees with multi word prep ositions as anchors Examples of these are Sr Sr
NP VPr NP VP
N VP NP
NA N V S1 NA
Esso V S1 N NA Esso said NPr VPr
said NPr VP Tuesday D NPf VP NP NA NA D NPf V NP NA
the Nr V NP N
the Nr started N
N Nf started N Tuesday N Nf production NA NA
Whiting field Whiting field production
Figure Time NP trees Two di erent attachments
Sr Sr
NP VPr NP S NA
N VP NP NA N NP VP Sr
it V NPr N today N V NPr
NP VP has NPf PP today r it has NPf PP NA NA
NPf NP V NPr D NPf P NPr D NPf P NPr NA NA NA
no N on D NPf no N on D NPf NA D NPf N claimed D NPf NA NA NA
bearing our Nr bearing our Nr
N Nf the N yesterday two N N Nf NA NA
work force fire lives work force
Figure Time NPs in di erent p ositions vxN nxN and Ns
ahead of contrary to at variance with and as recently as
CHAPTER MODIFIERS
Sr
αnx0Vnx1[announced]
NP VPr
αNXN[Michael] (1) βNvx[yesterday] (2) αNXN[program] (2.2) N NP VP NA βAn[late] (1.1) βDnx[a] (0) βNn[buy-back] (1)
Michael Nr V NPr
A Nf announced D NPf NA NA
late yesterday a Nr
N Nf NA
buy-back program
Figure Time NPs Derived tree and Derivation Nvx p osition
p osition and category mo di ed
pre sentential p ost NP p ost VP
Complementtyp e Smodi er NP mo di er VP mo di er
S complement Pss nxPs vxPs
NP complement Pnxs nxPnx vxPnx
no complement Ps nxP vxP
exhaustive
Table Prep osition Anchored Mo di ers
with Clove healthy the veterinarian s bill will b e more a ordable Pss
PP PP
The frisb ee in the brambles was hidden nxPnx
PP PP
Cloveplayed frisb ee outside vxP
PP PP
Cloveplayed frisb ee outside of the house vxPPnx
PP PP
Prep ositions that take NP complements assign accusative case to those complements see
section for details Most prep ositions take NP complements Sub ordinating conjunc
tions are analyzed in XTAG as Preps see Section for details Additionally a few non
conjunction prep ositions take S complements see Section for details
Clove healthy is an adjective small clause VPr
VP Sr NPr r NA VP* PP NA
PP Sf* NPf* PP P NP↓ VP* PP NA NA NA P1 P2
◊ ↓ ◊ ↓
P S P NP P◊ outside of
Pss nxPnx vxP vxPPnx
Figure Selected Prep ositional Phrase Mo di er trees Pss nxPnx vxP and vxPPnx
Adverbs
In the English XTAG grammar VP and S mo difying adverbs anchor the auxiliary trees ARBs
sARB vxARB and ARBvx allowing pre and p ost mo di cation of S s and VP s Besides
the VP and S mo difying adverbs the grammar includes adverbs that mo dify other categories
Examples of adverbs mo difying an adjective an adverb a PP an NP and a determiner are
shown in See Sections and for discussion of the puARBpuvx and
spuARB which handle pre verbal parenthetical adverbs and p ost sentential comma separated
adverbs
Mo difying an adjective
extremely go o d
rather tall
rich enough
Mo difying an adverb
o ddly enough
very well
Mo difying a PP
right through the wall
Mo difying a NP
quite some time
dverbs Because the letters in A Ad In the naming conventions for the XTAG trees ARB is used for a
and Adv are all used for other parts of sp eech a djective determiner and verb ARB was chosen to eliminate
ambiguity App endix D contains a full explanation of naming conventions
CHAPTER MODIFIERS
Mo difying a determiner
exactly vemen
XTAG has separate trees for each of the mo di ed categories and for pre and post mo di
cation where needed The kind of treatment given to adverbs here is very much in line with
the base generation approach prop osed by Ernst which assumes all p ositions where an
adverb can o ccur to b e base generated and that the semantics of the adverb sp eci es a range
of p ossible p ositions o ccupied by each adverb While the relevant semantic features of the
adverbs are not currently implemented implementation of semantic features is scheduled for
future work The trees for adverb anchored mo di ers are very similar in form to the adjec
tiveanchored mo di er trees Examples of two of the basic adverb mo di er trees are shown in Figure
S r inv : - wh : <1> displ-const : <2> agr : <3> VPr assign-case : <4> passive : <1> mode : <5> displ-const : <2> tense : <6> assign-case : <3> assign-comp : <7> comp : <8> nil assign-comp : <4> tense : <5> agr : <6> mode : <7>
Ad◊ wh : <1> S* inv : <1> NA displ-const : <2> agr : <3> ◊ assign-case : <4> VP* passive : <1> Ad mode : <5> NA displ-const : <2> tense : <6> assign-case : <3> assign-comp : <7> assign-comp : <4> comp : <8> tense : <5> sub-conj : nil agr : <6> sub-conj : nil mode : <7>
comp : nil
a b
Figure Adverb Trees for pre mo di cation of S ARBs a and p ost mo di cation of a
VP vxARB b
Like the adjectiveanchored trees these trees also have the NA constraint on the fo ot no de to
restrict the numb er of derivations pro duced for a sequence of adverbs Features of the mo di ed
category are passed from the fo ot no de to the ro ot no de re ecting correctly that these typ es of
prop erties are una ected by the adjunction of an adverb A summary of the categories mo di ed
and the p osition of adverbs is given in Table
Position with resp ect to item mo di ed
Category Mo di ed Pre Post
S ARBs sARB
VP ARBvx puARBpuvx vxARB
A ARBa aARB
PP ARBpx pxARB
ADV ARBarb arbARB
NP ARBnx
Det ARBd
Table Simple Adverb Anchored Mo di ers
In the English XTAG grammar no traces are p osited for wh adverbs in line with the
base generation approach Ernst for various p ositions of adverbs Since convincing
arguments have b een made against traces for adjuncts of other typ es e g Baltin and
since the reasons for wanting traces do not seem to apply to adjuncts we make the general
assumption in our grammar that adjuncts do not leave traces Sentence initial wh adverbs
select the same auxiliary tree used for other sentence initial adverbs ARBs with the feature
is as in wh Under this treatment the derived tree for the sentence How did you fal l
Figure with no trace for the adverb
There is one more adverb mo di er tree in the grammar which is not included in Table
This tree shown in Figure has a complex adverb phrase and is used for wh two adverb
phrases that occur sentence initially such as in sentence Since how is the only wh
adverb it is the only adverb that can anchor this tree
how quickly did Srini x the problem
Fo cus adverbs suchasonly even just and at least are also handled by the system Since the
syntax allows fo cus adverbs to app ear in practically any p osition these adverbs select most of
the trees listed in Table It is left up to the semantics or pragmatics to decide the correct
scop e for the fo cus adverb for a given instance In terms of the ability of the fo cus adverbs
to mo dify at di erent levels of a noun phrase the fo cus adverbs can mo dify either cardinal
determiners or noun cardinal noun phrases and cannot mo dify at the level of noun The tree
for adverbial mo di cation of noun phrases is in shown Figure a
In addition to at least the system handles the other two word adverbs at most and up
to and the three word as as adverb constructions where an adjective substitutes b etween the
two o ccurrences of as An example of a three word as as adverb is as little as Except for the
abilityof at least to mo dify many di erenttyp es of constituents as noted ab ove the multi word
adverbs are restricted to mo difying cardinal determiners Example sentences using the trees in
Figure are shown in
CHAPTER MODIFIERS
S r
Ad S NA
how V S NA
did NP VPr
N V VP NA
you ε V
fall
Figure Derived tree for How did you fal l Sr invlink : <3> + inv : <3> wh : <1> displ-const : <4> agr : <5> assign-case : <6> mode : <7> tense : <8> assign-comp : <9> comp : <10> nil
Adr wh : <1> S* inv : <3> wh : <2> + NA displ-const : <4> agr : <5> assign-case : <6> mode : <7> tense : <8> assign-comp : <9> comp : <10> sub-conj : nil sub-conj : nil comp : nil
Ad◊ wh : <2> Adc↓
Figure Complex adverb phrase mo di er ARBarbs
CHAPTER MODIFIERS
Fo cus Adverb mo difying an NP
only amemb er of our crazy family could pull o that kind of a stunt
even a ying saucer sighting would seem interesting in comparison
with your story
The rep ort includes a prop osal for at least a partial impasse in negotiations
Multi word adverbs mo difying cardinal determiners
at most ten p eople came to the party
They gave monetary gifts of as little as ve dollars
Dr Dr NA NA
NPr Ad Df * Ad1 Df * NA NA
◊
Ad NPf * P1 ◊ A↓ P2 ◊ P◊ Ad◊
ARBnx PaPd PARBd
a b c
Figure Selected Fo cus and Multi word Adverb Mo di er trees ARBnx PARBd and
PaPd
The grammar also includes auxiliary trees anchored by multi word adverbs like a little a
bit a mite sort of kind of etc
Multi word adverbs like sort of and kind of can pre mo dify almost any non clausal category
The only strict constraint on their o ccurrence is that they can t mo dify nouns in which case an
adjectival interpretation would obtain The category which they scop e over can be directly
determined from their p osition except for when they o ccur sentence nally in which case they
to mo dify VP s The complete list of auxiliary trees anchored by these adverbs are assumed
are as follows NPax NPpx NPnx NPvx vxNP NParb Selected trees are shown in
Figure and some examples are given in
John is sort of tired
AP
John is sort of to the right
PP
Note that there are semantic lexical constraints even for the categories that these adverbs can mo dify and
no doubt invite a more in depth analysis APr VPr VPr
Ad1 APf * Ad1 VPf * VPf * Ad1 NA NA NA NA NA NA
N◊ P◊ N◊ P◊ N◊ P◊
NPax NPvx vxNP
a b c
Figure Selected Multi word Adverb Mo di er trees for adverbs like sort of kind of
NPax NPvx vxNP
John could have b een sort of eating the cake
VP
John has b een eating his cake sort of slowly
AD V
There are some multi word adverbs that are however not so free in their distribution
Adverbs like a little a bit amitemo dify AP s in predicative constructions sentences with the
copula and small clauses AP complements in sentences with raising verbs and AP s when they
are sub categorized for by certain verbs e g John felt angry They can also p ost mo dify VP s
and PP s though not as freely as AP s Finally theyalso function as downtoners for almost
all adverbials Some examples are provided in
Mickey is a little tired
AP
The medicine has eased John s pain a little
VP
John is a little to the right
PP
John has b een reading his b o ok a little loudly
AD V
Following their b ehavior as describ ed ab ove the auxiliary trees they anchor are DAax
DApx vxDA DAarb DNax DNpx vxDN DNarb Selected trees are shown in
Figure
Lo cative Adverbial Phrases
Lo cative adverbial phrases are multi word adverbial mo di ers whose meanings relate to spatial
lo cation Lo catives consist of a lo cative adverb such as ahead or downstream preceded by
They can also app ear b efore NP s as in John wants a little sugar However here they function as
multi word determiners and should not b e analyzed as adverbs
It is to b e noted that this analysis whichallows these multiword adverbs to mo dify adjectival phrases as
well as adverbials will yield not necessarily desirable multiple derivations for a sentence like John is a little
unecessarily stupid In one derivation a little mo di es the AP and in the other case it mo di es the adverb
CHAPTER MODIFIERS
VPr APr PPr
VPf * Ad1 Ad1 APf * Ad1 PPf * NA NA NA NA NA NA
D◊ A◊ D◊ A◊ D◊ N◊
vxDA DAax DNpx
a b c
Figure Selected Multi word Adverb Mo di er trees for adverbs like a little a bit vxDA
DAax DNpx
an NP an adverb or nothing as in Examples resp ectively The mo di er as a
whole describ es a p osition relative to one previously sp eci ed in the discourse The nature of
the relation which is usually a direction is sp eci ed by the anchoring lo cativeadverb behind
east If an NP or a second adverb is present in the phrase it sp eci es the degree of the relation
for example three city blocks many meters and far
The accident three blocks ahead stopp ed tra c
The ship sank far o shor e
The trouble ahead distresses me
Lo catives can mo dify NPs VPs and Ss They mo dify NPs only by right adjoining p ost
p ositively as in Example Post p ositive is also the more common p osition when a lo cative
mo di es either of the other categories Lo catives pre mo dify VPs only when separated by
balanced punctuation commas or dashes The trees lo catives select when mo difying NPs are
shown in Figure
NP r NP r
NP 1* AdvP NP 1* AdvP NA NA
NP 2↓ Ad◊ Ad◊
Figure Lo cativeModi erTrees nxnxARB nxARB
When the lo cative phrase consists of only the anchoring lo cativeadverb as in Example
it uses the nxARB tree shown in Figure and its VP analogue vxARB In addition
these are the trees selected when the lo cative anchor is mo di ed byanadverb expressing degree
as in Example The degree adverb adjoins on to the anchor using the ARBarb tree which
is describ ed in Section Figure shows an example of these trees in action Though
there is a tree for a pre sentential lo cative phrase nxARBs there is no corresp onding p ost
sentential tree as it is highly debatable whether the p ost sentential version actually has the
entire sentence or just the preceding verb phrase as its scop e Thus in accordance with XTAG
practice which considers ambiguous p ost sentential mo di ers to be VP mo di ers rather than
S mo di ers there is only a vxnxARB tree as shown in Figure
S r S r
S AdvP NP VPr NA
N VP Adr NP VP NP r Ad NA
N V NP r D NP f back I V NP r Ad Ad1 NA NA NA
I left D NP f three N left D NP f far back NA NA
my N holes my N
toupee toupee
Figure Lo cative Phrases featuring NP and Adverb Degree Sp eci cations
One p ossible analysis of lo cative phrases with NPs mightmaintain that the NP is the head
with the lo cative adverb mo difying the NP This is initially attractive b ecause of the similarity
to time NPs which also feature NPs that can mo dify clauses This analysis seems insu cient
however in light of the fact that virtually any NP can o ccur in lo cative phrases as in example
Therefore in the XTAG analysis the lo cative adverb anchors the lo cative phrase trees
A complete summary of all trees selected bylocatives is contained in Table adverbs
select the lo cative trees
I left my toup ee and putter three holes back
Though nearly all of these adverbs are spatial in nature this numb er also includes a few temp oral adverbs
suchas ago that also select these trees
CHAPTER MODIFIERS
Degree Phrase Typ e
Category Mo di ed NP Ad None
NP nxnxARB nxARB
VP p ost vxnxARB vxARB
VP pre punct separated punxARBpuvx puARBpuvx
S nxARBs ARBs
Table Lo cative Mo di ers
Chapter
Auxiliaries
Although there has been some debate ab out the lexical category of auxiliaries the English
XTAG grammar follows McCawley Haegeman and others in classifying auxil
iaries as verbs The category of verbs can therefore be divided into two sets main or lexical
verbs and auxiliary verbs which can co o ccur in a verbal sequence Only the highest verb in a
verbal sequence is marked for tense and agreement regardless of whether it is a main or auxiliary
verb Some auxiliaries be do and have share with main verbs the prop erty of having overt
morphological marking for tense and agreement while the mo dal auxiliaries do not However
all auxiliary verbs di er from main v erbs in several crucial ways
Multiple auxiliaries can o ccur in a single sentence while a matrix sentence may have at
most one main verb
Auxiliary verbs cannot o ccur as the sole verb in the sentence but must b e followed bya
main verb
All auxiliaries precede the main verbinverbal sequences
Auxiliaries do not sub categorize for any arguments
Auxiliaries imp ose requirements on the morphological form of the verbs that immediately
follow them
Only auxiliary verbs invert in questions with the sole exception in American English of
main verb be
An auxiliary verb must precede sentential negation e g John not goes
Auxiliaries can form contractions with sub jects and negation e g he l l won t
The restrictions that an auxiliary verb imp oses on the succeeding verb limits the sequence of
verbs that can o ccur In English sequences of up to veverbs are allowed as in sentence
Some dialects particularly British English can also invert main verb have in yes no questions e g have
you any Grey Poupon This is usually attributed to the in uence of auxiliary have coupled with the historic
fact that English once allowed this movement for all verbs
CHAPTER AUXILIARIES
The music should have b een b eing played for the president
The required ordering of verb forms when all veverbs are present is
mo dal base p erfective progressive passive
The rightmost verb is the main verb of the sentence While a main verb sub categorizes for the
arguments that app ear in the sentence the auxiliary verbs select the particular morphological
forms of the verb to followeach of them The auxiliaries included in the English XTAG grammar
are listed in Table by typ e The third column of Table lists the verb forms that are
required to follow eachtyp e of auxiliary verb
TYPE LEX ITEMS SELECTS FOR
mo dals can could may might wil l base form
would ought shal l should e g wil l go might come
need
p erfective have past participle
e g has gone
progressive be gerund
e g is going was coming
passive be past participle
e g was helped by Jane
do supp ort do base form
e g did go does come
in nitiveto to base form
e g to go to come
Table Auxiliary Verb Prop erties
Non inverted sentences
This section and the sections that follow describ e how the English XTAG grammar accounts
for prop erties of the auxiliary system describ ed ab ove
In our grammar auxiliary trees are added to the main verb tree by adjunction Figure
shows the adjunction tree for non inverted sentences
The restrictions outlined in column of Table are implemented through the features
mo de p erfect progressive and passive The syntactic lexicon entries for the
auxiliaries gives values for these features on the fo ot no de VP in Figure Since the top
features of the fo ot no de must eventually unify with the b ottom features of the no de it adjoins
There are American dialects particularly in the South whichallow double mo dals suchasmight could and
might should These constructions are not allowed in the XTAG English grammar
Wesaw this tree brie y in section but with most of its features missing The full tree is presented
here VPr assign-comp : <1> neg : <2> agr : <3> mainv : <4> tense : <5> mode : <6> assign-case : <7> displ-const : set1 : <8> progressive : <9> perfect : <10> conditional : <11>
V◊ assign-comp : <1> VP* displ-const : set1 : - neg : <2> NA progressive : <9> agr : <3> perfect : <10> mainv : <4> conditional : <11> tense : <5> mode : <6> assign-case : <7>
displ-const : set1 : <8>
Figure Auxiliary verb tree for non inverted sentences Vvx
onto for the sentence to b e valid this enforces the restrictions made by the auxiliary no de In
addition to these feature values each auxiliary also gives values to the anchoring no de V to
b e passed up the tree to the ro ot VP VP no de there they will b ecome the new features for
r
the top VP no de of the sentential tree Another auxiliary maynow adjoin on top of it and so
forth These feature values thereby ensure the prop er auxiliary sequencing Figure shows
the auxiliary trees anchored by the four auxiliary verbs in sentence Figure shows
the nal tree created for this sentence
The general English restriction that matrix clauses must have tense or b e imp eratives is
enforced by requiring the top S no de of a sentencetohave mo de ind imp indicativeor
imp erative Since only the indicative and imp erativesentences have tense non tensed clauses
are restricted to o ccurring in emb edded environments
Noun verb contractions are lab eled NVC in their part of sp eech eld in the morphological
database and then undergo sp ecial pro cessing to split them apart into the noun and the reduced
verb b efore parsing The noun then selects its trees in the normal fashion The contraction say
CHAPTER AUXILIARIES
VP r VPr conditional : <1> conditional : <1> perfect : <2> progressive : <2> progressive : <3> displ-const : set1 : <3> displ-const : set1 : <4> assign-case : <5> assign-case : <4> mainv : <6> mode : <5> agr : <7> tense : <6> neg : <8> agr : <7> assign-comp : <9> neg : <8> tense : <10> pres assign-comp : <9> mode : <11> ind perfect : <10> + mainv : <11> -
V assign-comp : <9> VP* displ-const : set1 : - neg : <8> NA progressive : <3> agr : <7> perfect : <2> V VP* mainv : <6> conditional : <1> assign-comp : <9> displ-const : set1 : - tense : <10> mode : base neg : <8> NA progressive : <2> mode : <11> agr : <7> perfect : <10> assign-case : <5> mainv : <11> conditional : <1> displ-const : set1 : <4> tense : <6> mode : ppart displ-const : set1 : - mode : <5> passive : - assign-case : nom assign-case : <4> assign-comp : ind_nil/that/rel/if/whether displ-const : set1 : <3> mainv : - displ-const : tense : pres set1 : - mode : ind mode : base
should have
VPr VPr conditional : <1> conditional : <1> perfect : <2> perfect : <2> displ-const : set1 : <3> progressive : <3> assign-case : <4> displ-const : set1 : <4> mode : <5> assign-case : <5> tense : <6> mode : <6> agr : <7> tense : <7> neg : <8> agr : <8> assign-comp : <9> neg : <9> progressive : <10> + assign-comp : <10> mainv : <11> - mainv : <11> -
V assign-comp : <9> VP* displ-const : set1 : - V assign-comp : <10> VP* displ-const : set1 : - neg : <8> NA progressive : <10> neg : <9> NA progressive : <3> agr : <7> perfect : <2> agr : <8> perfect : <2> mainv : <11> conditional : <1> mainv : <11> conditional : <1> tense : <6> mode : ger tense : <7> mode : ppart mode : <5> mode : <6> passive : + assign-case : <4> assign-case : <5> mainv : + displ-const : set1 : <3> displ-const : set1 : <4> displ-const : set1 : - displ-const : set1 : - weak : - assign-case : none mode : ppart mode : ger
been being
Figure Auxiliary trees for The music should have been being played Sr
NP VPr
D N V VP NA
the music should V VP NA
have V VP NA
been V VP NA
being V
played
Figure The music should have been being played
l l or d likewise selects the normal auxiliary verb tree Vvx However since the contracted
form rather than the verb stem is given in the morphology the contracted form must also b e
listed as a separate syntactic entry These entries have all the same features of the full form
of the auxiliary verbs with tense constraints coming from the morphological entry e g it s is
listed as it s NVC sg PRES The ambiguous contractions d had would and s has is
multiple entries for those lexical b ehave like other ambiguous lexical items there are simply
items in the lexicon each with di erent features In the resulting parse the contracted form is
shown with features appropriate to the full auxiliary it represents
Inverted Sentences
In inverted sentences the two trees shown in Figure adjoin to an S tree anchored by a
main verb The tree in Figure a is anchored by the auxiliary verb and adjoins to the S
no de while the tree in Figure b is anchored byanempty element and adjoins at the VP
CHAPTER AUXILIARIES
no de Figure shows these trees anchored by wil l adjoined to the declarative transitive
tree anchored by main verb buy
VPr conditional : <1> perfect : <2> progressive : <3> Sr displ-const : set1 : - displ-const : set1 : <7> assign-case : <8> neg : <1> tense : <9> agr : <2> neg : <10> tense : <3> assign-comp : <11> progressive : <7> agr : <4> perfect : <8> passive : <5> conditional : <9> mode : <6> mainv : <12> - assign-case : <5> mode : <4> inv : +
V assign-comp : <11> VP* conditional : <1> neg : <10> NA perfect : <2> agr : <4> progressive : <3> V◊ neg : <1> S* progressive : <7> mainv : <12> displ-const : set1 : - agr : <2> NA perfect : <8> tense : <9> agr : <4> tense : <3> conditional : <9> mode : <6> passive : <5> mode : <4> ind assign-case : <5> assign-case : <8> mode : <6> assign-case : <5> agr : <6> displ-const : set1 : <7> agr : <6> displ-const : set1 : + displ-const : set1 : + comp : nil inv : -
ε
a b
Figure Trees for auxiliary verb inversion Vs a and Vvx b
The feature displ const ensures that both of the trees in Figure must adjoin to
an elementary tree whenever one of them do es For more discussion on this mechanism which
simulates tree lo cal multi comp onent adjunction see Ho ckey and Srinivas The tree in
Figure b anchored by represents the originating p osition of the inverted auxiliary Its
adjunction blo cks the assign case values of the VP it dominates from b eing co indexed
with the case value of the sub ject Since assign case values from the VP are blo cked
only be co indexed with the assign case value of the case value of the sub ject can
the inverted auxiliary Figure a Consequently the inverted auxiliary functions as the
case assigner for the sub ject in these inverted structures This is in contrast to the situation in
uninverted structures where the anchor of the highest leftmost VP assigns case to the sub ject
The declarative transitivetreewas seen in section Sr
V S NA
will NP VPr
N V VP NA
John ε V NP
buy D N
a backpack
Figure wil l John buy a backpack
see section for more on case assignment The XTAG analysis is similar to GB accounts
where the inverted auxiliary plus the anchored tree are taken as representing I to C movement
Do Supp ort
It is well known that English requires a mechanism called do supp ort for negated sentences
and for inverted yes no questions without auxiliaries
John do es not want a car
John not wants a car
John will not want a car
Do you wanttoleavehome
wantyou to leave home
will you wanttoleavehome
CHAPTER AUXILIARIES
In negated sentences
The GB analysis of do supp ort in negated sentences hinges on the separation of the INFL and
VP no des see Chomsky Jackendo and Chomsky The claim is that
the presence of the negative morpheme blo cks the main verb from getting tense from the INFL
no de thereby forcing the addition of a verbal lexeme to carry the in ectional elements If
an auxiliary verb is present then it carries tense but if not p eriphrastic or dummy do is
required This seems to indicate that do and other auxiliary verbs would not co o ccur and
indeed this is the case see sentences Auxiliary do is allowed in English when no
negative morpheme is present but this usage is marked as emphatic Emphatic do is also not
allowed to co o ccur with auxiliary verbs sentences
Wewillhavedobought a sleeping bag
Wedowillhavebought a sleeping bag
You do have a backpack don t you
I do want togo
You do can have a backpack don t you
I did havehadabackpack
At present the XTAG grammar do es not have analyses for emphatic do
In the XTAG grammar do is prevented from co o ccurring with other auxiliary verbs by a
requirement that it adjoin only onto main verbs mainv It has indicative mo de so
no other auxiliaries can adjoin ab ove it The lexical item not is only allowed to adjoin onto a
non indicative and therefore non tensed verb Since all matrix clauses must b e indicative or
imp erative a negated sentence will fail unless an auxiliary verb either do or another auxiliary
adjoins somewhere ab ove the negative morpheme not In addition to forcing adjunction of an
auxiliary this analysis of not allows it freedom to move around in the auxiliaries as seen in the
sentences
John will havehadabackpack
John not will have had a backpack
John will not have hadabackpack
John will have nothadabackpack
Earlier we said that indicative mo de carries tense with it Since only the topmost auxiliary carries the tense
any subsequentverbs must not have indicativemode
In inverted yes no questions
In inverted yes no questions do is required if there is no auxiliary verb to invert as seen in
sentences replicated here as
do you wanttoleavehome
wantyou to leave home
will you wanttoleavehome
do you will want to leave home
In English unlike other Germanic languages the main verb cannot move to the b eginning
of a clause with the exception of main verb be In a GB accountofinverted yes no questions
the tense feature is said to b e in C at the front of the sentence Since main verbs cannot move
they cannot pick up the tense feature and do supp ort is again required if there is no auxiliary
verb to p erform the role Sentence shows that do do es not interact with other auxiliary
verbs even when in the inverted p osition
In XTAG trees anchored by a main verb that lacks tense are required to have an auxiliary
verb adjoin onto them whether at the VP no de to form a declarativesentence or at the S no de
to form an inverted question Do selects the inverted auxiliary trees given in Figure just
as other auxiliaries do so it is available to adjoin onto a tree at the S no de to form a yes no
question The mechanism describ ed in section prohibits do from co o ccurring with other
auxiliary verbs even in the inverted p osition
In nitives
The in nitive to is considered an auxiliary verb in the XTAG system and selects the auxiliary
tree in Figure To like do doesnotinteract with the other auxiliary verbs adjoining only
to main verb base forms and carrying in nitive mo de It is used in embedded clauses both
with and without a complementizer as in sentences Since it cannot be inverted
it simply do es not select the trees in Figure
John wants to have a backpack
John wants Mary to haveabackpack
John wants for Mary to have a backpack
The usage of in nitival to interacts closely with the distribution of null sub jects PRO and
is describ ed in more detail in section
The inversion of main verb have in British English was previously noted
CHAPTER AUXILIARIES
Semi Auxiliaries
Under the category of semi auxiliaries wehave placed several verbs that do not seem to closely
follow the b ehavior of auxiliaries One of these auxiliaries dare mainly behaves as a mo dal
and selects for the base form of the verb The other semi auxiliaries all select for the in nitival
form of the verb Examples of this second typ e of semi auxiliary are used to ought to get to
have to andBE to
Marginal Mo dal dare
The auxiliary dare is unique among mo dals in that it both allows DO supp ort and exhibits
a past tense form It clearly falls in mo dal p osition since no other auxiliary except do may
precede it in linear order Examples app ear b elow
she dare not have b een seen
shedoesnotdare succeed
Jerry dared not lo ok left or right
only mo dels dare wear such extraordinary out ts
dare Dale tell her the secret
Louise had dared not tell a soul
As mentioned ab ove auxiliaries are prevented from having DO supp ort within the XTAG
system To allow for DO supp ort in this case we had to create a lexical entry for dare that
allowed it to have the feature mainv and to have base mo de this measure is what also
allows dar e to o ccur in double mo dal sequences A second lexical entry was added to handle
the regular mo dal o ccurrence of dare Additionally all other mo dals are classi ed as b eing
present tense while dare has b oth present and past forms To handle this b ehavior dare was
given similar features to the other mo dals in the morphology minus the sp eci cation for tense
Other semi auxiliaries
The other semi auxiliaries all select for the in nitival form of the verb Many of these auxiliaries
allow for DO supp ort and can app ear in both base and past participle forms in addition to
b eing able to stand alone indicativemode Examples of this typ e app ear b elow
Alex used to attend karate workshops
Angelina mighthave used to b elieveinfate
ttobeaphysical therapist Rich did not used to wan
Some sp eakers accept dare preceded by a mo dal as in I might dare nish this report today In the XTAG
analysis this particular double mo dal usage is accounted for Other cases of double mo dal o ccurrence exist in
some dialects of American English although these are not accounted for in the system as was mentioned earlier
Mickmightnot have to play the game tonight
Singer had to have b een there
Heather has got to nish that pro ject b efore she go es insane
The auxiliaries ought to and BE to may not b e preceded byany other auxiliary
Bi ought to have b een working harder
Carson do es ought to havebeenworking harder
the party is to take place this evening
the partyhadbeen to take place this evening
The trickiest element in this group of auxiliaries is used to While the other verbs behave
according to standard in ection for auxiliaries usedtohas the same form whether it is in mo de
base past participle or indicative forms The only connection used to maintains with the
in nitival form use is that o ccasionally the bare form use will app ear with DO supp ort Since
the three mo des mentioned ab ove are mutually exclusive in terms of b oth the morphology and
the lexicon used has three entries in each
Other Issues
There is a lingering problem with the auxiliaries that stems from the fact that there currently
is no way to distinguish b etween the main verb and auxiliary verb b ehaviors for a given letter
string within the morphology This situation results in many unacceptable sentences b eing
successfully parsed by the system Examples of the unacceptable sentences are given b elow
the miller cans tell a go o d story vs the farmer cans peachesinJuly
David wills have nished bynoon vs the old man wills his fortune to me
Sarah needs not leave vs Sarah needs to leave
Jennifer dares not b e seen vs the young woman dares him to do the stun t
Lila do es use to like b eans vs Lila do es use her new co okware
Chapter
Conjunction
Intro duction
The XTAG system can handle sentences with conjunction of two constituents of the same
syntactic category The co ordinating conjunctions which select the conjunction trees are and
or and but There are also multi word conjunction trees anchored by either or neither nor
both and and as wel l as There are eightsyntactic categories that can be co ordinated and in
each case an auxiliary tree is used to implement the conjunction These eight categories can
be considered as four di erent cases as describ ed in the following sections In all cases the
two constituents are required to b e of the same syntactic category but there may also b e some
additional constraints as describ ed b elow
Adjective Adverb Prep osition and PP Conjunction
Each of these four categories has an auxiliary tree that is used for conjunction of two constituents
of that category The auxiliary tree adjoins into the left hand side comp onent and the right
hand side comp onent substitutes into the auxiliary tree
Figure a shows the auxiliary tree for adjective conjunction and is used for example
in the derivation of the parse tree for the noun phrase the dark and dreary day as shown in
Figure b The auxiliary tree adjoins onto the no de for the left adjective and the right
e substitutes into the right hand side no de of the auxiliary tree The analysis for adverb adjectiv
prep osition and PP conjunction is exactly the same and there is a corresp onding auxiliary tree
for each of these that is identical to that of Figure a except of course for the no de lab els
Noun Phrase and Noun Conjunction
The tree for NP conjunction shown in Figure a has the same basic analysis as in the
previous section except that the wh and case features are used to force the two noun
phrases to have the same wh and case values This allows for example he and she
Agreement is wrote the book together while disallowing he and her wrote the book together
We b elieve that the restriction of but to conjoining only two items is a pragmatic one and our grammars
accepts sequences of anynumb er of elements conjoined by but NP
DetP Nr
D A Nf NA
A the A1 Conj A day
A1* Conj◊ A2↓ dark and dreary
a b
Figure Tree for adjective conjunction a CONJa and a resulting parse tree
lexicalized since the various conjunctions b ehave di erently With and the ro ot agr num
value is plural no matter what the numb er of the two conjuncts With or however the ro ot
agr num is co indexed with the agr num feature of the right conjunct This ensures
that the entire conjunct will b ear the numb er of b oth conjuncts if they agree Figure b
or of the most recent one if they di er Either the boys or John is going to help you There
is no rule p er se on what the agreement should b e here but p eople tend to make the verb agree
with the last conjunct cf Quirk et al section for discussion The tree for N
for the no de lab els The multi word conjunction is identical to that for the NP tree except
conjunctions do not select the N conjunction tree the both dogs and cats
Determiner Conjunction
In determiner co ordination all of the determiner feature values are taken from the left deter
miner and the only requirementisthatthe wh feature is the same while the other features
such as card are unconstrained For example which and what and al l but one are both
acceptable determiner conjunctions but which and al l is not
how many and which p eople camp frequently
some or which p eople enjoy nature
Sentential Conjunction
The tree for sentential conjunction shown in Figure is based on the same analysis as the
features The mo de conjunctions in the previous two sections with a slight di erence in
CHAPTER CONJUNCTION
NP NP case : <1> nom/acc decreas : <4> wh : <2> - card : <5> displ-const : <3> quan : <6> conj : <4> or definite : <7> const : <5> - gen : <8> gen : <6> - const : <9> conj : <1> definite : <7> - displ-const : <10> quan : <8> - wh : <2> card : <9> - case : <3> decreas : <10> - predet : <11> agr : num : <12> plur
Conj1 ◊ conj : <1> NP1 * decreas : <4> Conj2 ◊ NP2 ↓ wh : <2> NA card : <5> case : <3> nom/acc quan : <6> definite : <7> gen : <8> Conj1 NP1 Conj2 NP const : <9> NA displ-const : <10> wh : <2> case : <3> either N or N
aardvarks emus
a b
Figure Tree for NP conjunction CONJnx CONJnx and a resulting parse tree
feature is used to constrain the two sentences b eing conjoined to havethesamemodesothatthe
day is dark and the phone never rang is acceptable but the day dark and the phone never rang
is not Similarly the two sentences must agree in their wh comp and extracted
features Co indexation of the comp feature ensures that either both conjuncts have the
same complementizer or there is a single complementizer adjoined to the complete conjoined
S The assign comp feature feature is used to allow conjunction of in nitival sentences
and to sleep is a good life suchas to read
Comma as a conjunction
We treat comma as a conjunction in conjoined lists It anchors the same trees as the lexical
conjunctions but is considerably more restricted in how it combines with them The trees
anchored by commas are prohibited from adjoining to anything but another comma conjoined
element or a non co ordinate element All scop e p ossibilities are allowed for elements co ordi
nated with lexical conjunctions Thus structures such as Tree a are p ermitted with
each element stacking sequentially on top of the rst element of the conjunct while structures
suchasTree b are blo cked
See section for an explanation of the mo de feature
See section for an explanation of the assign comp feature D conj : <1> wh : <2> const : <3> decreas : <4> gen : <5> card : <6> quan : <7> agr : <8> definite : <9> predet : <10> displ-const : <11>
D1 * wh : <2> Conj◊ conj : <1> D2 ↓ wh : <2> NA const : <3> decreas : <4> gen : <5> card : <6> quan : <7> agr : <8> definite : <9> predet : <10>
displ-const : <11>
Figure Tree for determiner conjunction d CONJd ps
CHAPTER CONJUNCTION
S comp : <2> extracted : <3> wh : <4> conj : <1> displ-const : <5> mode : <6> ind/inf/ger/nom/prep/imp
S1* comp : <2> Conj◊ conj : <1> S2↓ comp : <2> NA extracted : <3> extracted : <3> wh : <4> wh : <4> displ-const : <5> assign-comp : inf_nil/ind_nil assign-comp : inf_nil/ind_nil mode : <6>
mode : <6>
Figure Tree for sentential conjunction s CONJs
NP NP
Nr Nr
A Nf A Nf NA NA
A1 Conj A roses A1 Conj A roses NA NA
A1 Conj A , fragrant beautiful , A1 Conj A NA NA
beautiful , red red , fragrant
a Valid tree with comma conjunction b Invalid tree
Figure
This is accomplished by using the conj feature which has the values and or but
and comma to di erentiate the lexical conjunctions from commas The conj values for a
comma anchored tree and and anchored tree are shown in Figure The feature conj
comma none on A in a only allows comma conjoined or non conjoined elements as the
left adjunct and conj none on A in a allows only a non conjoined element as the right
conjunct We also need the feature conj and or but none on the right conjunct of
the trees anchored by lexical conjunctions like b to blo ck comma conjoined elements from
substituting there Without this restriction we would get multiple parses of the NP in Tree
with the restrictions weonlyget the derivation with the correct scoping shown as a
Since comma conjoined lists can app ear without a lexical conjunction b etween the nal two
elements as shown in example we cannot force all comma conjoined sequences to end
with a lexical conjunction
So it is to o with many other spirits which we all know the spirit of Nazism or Com
munism scho ol spirit the spirit of a street corner gang or a fo otball team the spirit of
Rotary or the Ku Klux Klan Brown cd
Nr Nr
A neg : - N * f A neg : - Nf* conj : <14> comma NA conj : <14> and NA displ-const : <15> displ-const : <15>
A1 wh : - Conj conj : <14> A wh : - A wh : - Conj conj : <14> A wh : - NA displ-const : <15> neg : - 1 neg : - conj : none NA neg : - neg : - conj : comma/none displ-const : <15> conj : and/or/but/none
large , white red and white
Figure a CONJa a anchored by comma and b anchored by and
But not not but and not and not
We are analyzing conjoined structures suchas The women but not the men with a multi anchor
conjunction tree anchored bythe conjunction plus the adverb not The alternativeistoallow
not to adjoin to any constituent However this is the only construction where not can freely
occur onto a constituent other than a VP or adjective cf NEGvx and NEGa trees It can
also adjoin to some determiners as discussed in Section We want to allow sentences like
is shown in Figure and rule out those like The tree for the go o d example
There are similar trees for and not and not where is interpretable as either and or but and
a tree with not on the rst conjunct for not but
Beth grows basil in the house but not in the garden
Beth grows basil but not in the garden
CHAPTER CONJUNCTION
PP
PP1* Conj PP0 NA NA
but Ad PP2↓
not
Figure Tree for conjunction with but not px CONJARBpx
Although these constructions sound a bit odd when the two conjuncts do not have the
same numb er they are sometimes p ossible The agreement information for such NPs is always
that of the non negated conjunct his sons and not Bil l are in charge of doing the laundry
or not Bil l but his sons are in charge of doing the laundry Some p eople insist on having the
commas here but they are frequently absent in corpus data The agreement feature from the
non negated conjunct in passed to the ro ot NP asshown in Figure Aside from agreement
these constructions b ehave just like their non negated counterparts
To as a Conjunction
To can b e used as a conjunction for adjectives Fig and determiners when they denote
p oints on a scale
two to three degrees
high to very high temp eratures
As far as we can tell when the conjuncts are determiners they must b e cardinal
Predicative Co ordination
This section describ es the metho d for predicative co ordination including VP co ordination of
various kinds used in XTAG The description is derived from work describ ed in Sarkar and
Joshi It is imp ortant to say that this implementation of predicative co ordination is
not part of the XTAG release at the moment due massive parsing ambiguities This is partly
b ecause of the current implementation and also the inherentambiguities due to VP co ordination
that cause a combinatorial explosion for the parser We are trying to remedy b oth of these
which will be included as limitations using a probability mo del for co ordination attachments
part of a later XTAG release NP neg : + wh : <1> case : <2> nom/acc displ-const : <3> conj : <4> but agr : num : <5>
NP0 displ-const : <3> Conj conj : <4> NP2↓ wh : <1> NA case : <2> case : <2> wh : <1> agr : num : <5> conj : and/or/but/none
Ad NP1* displ-const : <3> but NA wh : <1> case : <2>
not
Figure Tree for conjunction with not but ARBnx CONJnx
This extended domain of lo cality in a lexicalized Tree Adjoining Grammar causes problems
when we consider the co ordination of such predicates Consider for instance the NP the
beansthatIbought from Alice in the Right No de Raising RNR construction has to b e shared
by the two elementary trees which are anchored by cooked and ate resp ectively
Harry co oked and Mary ate the b eans that I b ought from Alice
We use the standard notion of co ordination which is shown in Figure which maps two
constituents of like type but with di erentinterpretations into a constituent of the same typ e
We add a new op eration to the LTAG formalism in addition to substitution and adjunction
called conjoin later we discuss an alternative which replaces this op eration by the traditional
op erations of substitution and adjunction While substitution and adjunction take two trees
to give a derived tree conjoin takes three trees and comp oses them to give a derived tree
One of the trees is always the tree obtained by sp ecializing the schema in Figure for a
particular category The tree obtained will b e a lexicalized tree with the lexical anchor as the
conjunction and but etc
The conjoin op eration then creates a contraction between no des in the contraction sets of
the trees b eing co ordinated The term contraction is taken from the graph theoretic notion of
edge contraction In a graph when an edge joining two vertices is contracted the no des are
merged and the new vertex retains edges to the union of the neighb ors of the merged vertices
The conjoin op eration supplies a new edge b etween each corresp onding no de in the contraction
set and then contracts that edge
CHAPTER CONJUNCTION
NP
Nr
A Nf NA
A1 Conj Ar temperatures NA
high to Ad A NA
very high
Figure Example of conjunction with to
X
XXConj
Figure Co ordination schema
For example applying conjoin to the trees Conj and eats and dr ink s gives us the
derivation tree and derived structure for the constituent in shown in Figure
eats co okies and drinks b eer
Another way of viewing the conjoin op eration is as the construction of an auxiliary structure
from an elementary tree For example from the elementary tree dr ink s the conjoin op er
ation would create the auxiliary structure dr ink s shown in Figure The adjunction
op eration would now b e resp onsible for creating contractions b etween no des in the contraction
sets of the two trees supplied to it Such an approach is attractive for two reasons First it
the traditional op erations of substitution and adjunction Secondly it treats conj uses only
X as a kind of mo di er on the left conjunct X This approach reduces some of the parsing
ambiguities intro duced by the predicative co ordination trees and forms the basis of the XTAG
implementation
More information ab out predicative co ordination can b e found in Sarkar and Joshi
including an extension to handle gapping constructions Conj(and) S S VP VP
NP VP and VP VP and VP V NP Conj(and) VNP 3 1 eats cookies drinks beer
α (eats) α (drinks) Derived structure {1} {1} 2.2 2.2 α (cookies) α (beer)
Derivation tree
Figure An example of the conjoin op eration f g denotes a shared dep endency
α β (drinks) (eats) {1} {1} S VP S VP* and NP VP NP VP V NP V NP
drinks eats
S S VP α (eats) {1} VP NP and VP
V NP V NP β (drinks) {1}
eats drinks Derivation tree
John eats cookies and drinks beer
Figure Co ordination as adjunction
CHAPTER CONJUNCTION
Pseudo co ordination
The XTAG grammar do es handle one sort of verb pseudo co ordination Semi idiomatic phrases
such as try and and up and as in they might try and come to day are handled as multi
anchor mo di ers rather than as true co ordination These items adjoin to a Vnode using the
VCONJv tree This tree adjoins only to verbs in their base morphological non in ected form
The verb anchorofthe VCONJv must also b e in its base form as shown in examples
This blo cks rd p erson singular derivations which are the only p erson morphologically
marked in the present except when an auxiliary verb is presentortheverb is in the in nitive
He tried and came yesterday
They try and exercise three times a week
He wants to try and sell the puppies
Chapter
Comparatives
Intro duction
Comparatives in English can manifest themselves in many ways acting on many di erent
grammatical categories and often involving ellipsis A distinction must b e made at the outset
between twovery di erent sorts of comparatives those whichmake a comparison b etween two
prop ositions and those which compare the extent to which an entity has one prop erty to a
greater or lesser extent than another prop erty The former which we will refer to as proposi
tional comparatives is exempli ed in while the latter whic hwe will call metalinguistic
comparatives following Hellan is seen in
Ronaldo is more angry than Romario
Ronaldo is more angry than upset
In the extent to which Ronaldo is angry is greater than the extent to which Romario is
angry Sentence indicates that the extent to which Ronaldo is angry is greater than the
extenttowhich he is upset
Apart from certain of the elliptical cases b oth kinds of comparatives can b e handled straight
forwardly in the XTAG system Elliptical cases which are not presently covered include those
exempli ed by the following sentences which would presumably b e handled in the same way
as other sorts of VP ellipsis would
Ronaldo is more angry than Romario is
Bill eats more bro ccoli than George eats
Bill eats more bro ccoli than George do es
We turn to the analysis of metalinguistic comparatives rst
Metalinguistic Comparatives
A metalinguistic comparison can b e p erformed on basically all of the predicational categories
adjectives verb phrases prep ositional phrases and nouns as in the following examples
CHAPTER COMPARATIVES
The table is more long than wide AP
Clark more makes the rules than follows them VP
Calvin is more in the living ro om than in the kitchen PP
That unindenti ed amphibian in the bush is more frog than toad I would say NP
At present we only deal with the adjectival metalinguistic comparatives as in The
analysis given here for these can b e easily extended to prep ositional phrases and nominal com
paratives of the metalinguistic sort but as with co ordination in XTAG verb phrases will prove
more di cult
Adjectival comparatives app ear to distribute with simple adjectives as in the following
examples
Herb ert is more livid than angry
Herb ert is more livid and furious than angry
The more innovative than conventional medication cured everyone in the sickward
The elephant more wobbly than steady fell from the circus ball
Ar
Ad◊ Af* PP
P◊ A↓
Figure Tree for Metalinguistic Adjective Comparative ARBaPa
This patterning indicates that we can give these comparatives a tree that adjoins quite
freely onto adjectives as in Figure This tree is anchored by more less than To avoid
grammatically incorrect comparisons such as more brighter than dark the feature compar
is used to blo ck this tree from adjoining onto morphologically comparative adjectives The
fo ot no de is compar while brighter and its comparative siblings are compar We also
wish to blo ck strings like more brightest than dark which is accomplished with the feature
sup er indicating sup erlatives This feature is negative at the fo ot no de so that ARBaPa
cannot adjoin to sup erlatives like nicest whichare sp eci ed as sup er from the morphology
Furthermore the ro ot no de is sup er so that ARBaPa cannot adjoin onto itself and pro duce
monstrosities such as
The analysis given later for adjectival prop ositional comparatives pro duces aggregated compar adjectives
suchas more bright which will also b e incompatible as desired with ARBaPa
Herb ert is more less livid than angry than furious
Thus the use of the sup er feature is less to indicate sup erlativeness sp eci cally but rather to
indicate that the subtree b elowa sup er no de contains a full eshed comparison In the case
of lexical sup erlatives the comparison is against everything implicitly
A b ene t of the multiple anchor approach here is that we will never allow sentences such
as which would be p ermissible if we split the comparative comp onent and the than
comp onent of metalinguistic comparatives into two separate trees
Ronaldo is angrier than upset
We also see another variety of adjectival comparatives of the form more less than X which
indicates some prop erty which is more or less extreme than the prop erty X In a sentence such
as some prop erty is b eing said to hold of Francis such that it is of a kind with stupid and
that it exceeds stupid on some scale intelligence for example Quirk et al also note that these
constructions remark on the inadequacy of the lexical item Thus in it could be that
stupid is a starting p oint from which the sp eaker makes an approximation for some prop erty
which the sp eaker feels is beyond the range of the English lexicon but which expresses the
supreme lackofintellect of the individual it is predicated of
Francis is more than stupid
Romario is more than just upset
Taking our inspiration from ARBaPa we can handle these comparatives whichhave the
same distribution but contain an empty adjective by using the tree shown in Figure
Ar
Ad◊ A PP NA
ε P◊ Af*
Figure Tree for Adjective Extreme Comparative ARBPa
This sort of metalinguistic comparative also occurs with the verb phrase prep ositional
phrase and noun varieties
Clark more than makes the rules VP
Calvin s hands are more than near the co okie jar PP
That stu on her face is more than mud NP
Presumably the analysis for these would parallel that for adjectives though it has not yet b een
implemented
CHAPTER COMPARATIVES
Prop ositional Comparatives
Nominal Comparatives
Nominal comparatives are considered here to be those which compare the cardinality of two
sets of entities denoted by nominal phrases The following data lay out a basic distribution of
these comparatives
More vikings than mongols eat spam
More the vikings than mongols eat spam
Vikings eat less spaghetti than spam
More men that walk to the store than women who despise spam enjoyed the fo otball
game
More men than James likescotch on the ro cks
Elmer knows fewer martians than rabbits
Lo oking at these examples we are tempted to pro duce a tree for this construction that is
similar to ARBaPa However it is quite common for the than p ortion of these comparatives
to b e left out as in the following sentences
More vikings eat spam
Mongols eat less spam
Furthermore than NP cannot o ccur without more These facts indicate that we can and should
build up nominal comparatives with two separate trees The rst which allows a comparative
adverb to adjoin to a noun is given in Figure a The second is the noun phrase mo difying
prep ositional tree The tree CARBn is anchored by more less fewer and CnxPnx is anchored
by than The feature compar is used to ensure that only one CARBn tree can adjoin to any
given noun its fo ot no de is compar and the ro ot no de is compar All nouns are compar
and the compar value is passed up through all trees which adjoin to N or NP In order to ensure
that we do not allow sentences like Vikings than mongols eat spam the compar feature is
compar thus CnxPnx will adjoin only to NP s used The NP fo ot no de of CnxPnx is
whichhave b een already mo di ed by CARBn and thereby comparativized In this way we
capture sentences like en route to deriving sentences like in a principled and simple
manner
Further evidence for this approach comes from comparative clauses which are missing the
noun phrase which is b eing compared against something as in the following
The vikings ate more
We ignore here the interpretation in which the comparison covers the eating event fo cussing only on the one
which the comparison involves the stu b eing eaten Nr equiv : <1> super : - compar : + NP r
NP f* PP Ad◊ equiv : <1> Nf* compar : - compar : + super : - super : -
P◊ NP ↓
a CARBn tree b CnxPnx tree
Figure Nominal comparative trees
CHAPTER COMPARATIVES
The vikings ate more than a b oar
Sometimes the missing noun refers to an entityorsetavailable in the prior discourse while at
other times it is a reference to some anonymous unsp eci ed set The former is exempli ed in
a mini discourse such as the following
Calvin The mongols ate spam
Hobb es The vikings ate more
The latter can b e seen in the following example
Calvin The vikings ate a a b oar
Hobb es Indeed But in fact the vikings ate more than a b oar
Since the lone comparatives more less fewer have the same basic distribution as noun
phrases the tree in Figure is employed to capture this fact The ro ot no de of CARB is
compar Not only do es this accord with our intuitions ab out what the compar feature is
supp osed to indicate it also p ermits nxPnx to adjoin giving us strings suchas more than NP for free
NP
Ad◊ N NA
ε
Figure Tree for Lone Comparatives CARB
Thus by splitting nominal comparatives into multiple trees we make correct predictions
ab out their distribution with a minimal numb er of simple trees Furthermore wenow also get
certain comparative co ordinations for free once we place the requirement that nouns and noun
phrases must matchforcompar if they are to b e co ordinated This yields strings suchasthe
following
Julius eats more grap es and fewer b oars than avo cados
Were there more or less than fty p eople at the party
This sentence di ers from the metalinguistic comparison That stu on her faceismore than mud in that
it involves a comment on the quantity and or typ e of the compared NP whereas the other expresses that the
prop erty denoted by the compared noun is an inadequate characterization of the thing b eing describ ed
The structures are given in Figure Also it will blo ck strings like more men and women than
children under the imp ossible interpretation that there are more men than children but the
comparison of the quantityofwomen to children is not p erformed Unfortunately it will p ermit
comparative clauses such as more grapes and fewer than avocados under the interpretation in
which there are more grap es than avo cados and fewer of some unsp eci ed thing than avo cados see Figure
NP r
NP PP
NP 1 Conj NP P NP NA
Nr and Nr than N
Ad Nf Ad Nf avocados
more grapes fewer boars
NP r
NP f PP
NP 1 Conj NP P NP NA
Ad N or Ad N than Nr NA NA
more ε less ε N Nf NA
fifty people
Figure Comparative conjunctions
One asp ect of this analysis is that it handles the elliptical comparatives such as the following
Arnold kills more bad guys than Steven
CHAPTER COMPARATIVES
NP r
NP f PP NA
NP 1 Conj NP P NP NA
Nr and Ad N than N NA
Ad Nf fewer ε avocados NA
more grapes
Figure Comparative conjunctions
In a sense this is actually only simulating the ellipsis of these constructions indirectly However
consider the following sentences
Arnold kills more bad guys than I do
Arnold kills more bad guys than I
Arnold kills more bad guys than me
The rst of these has a pro verb phrase which has a nominative sub ject If we totally drop the
second verb phrase we nd that the second NP can b e in either the nominative or the accusative
case Prescriptive grammars disallow accusative case but it actually is more common to nd
accusative case use of the nominativeinconversation tends to sound rather sti and unnatural
This accords with the present analysis in which the second noun phrase in these comparatives
is the complement of than in nxPnx and receives its case marking from than This do es
mean that the grammar will not currently accept and indeed such sentences will only
et the fact that most sp eakers be covered by an analysis which really deals with the ellipsis Y
pro duce indicates that some sort of restructuring has o ccured that results in the kind of
structure the present analysis o ers
There is yet another distributional fact which falls out of this analysis When comparative
or comparativized adjectives mo dify a noun phrase they can stand alone or o ccur with a than
phrase furthermore they are obligatory when a than phrase is present
Hobb es is a b etter teacher
Hobb es is a b etter teacher than Bill
A more exquisite horse launched onto the racetrack
A more exquisite horse than Black Beauty launched onto the racetrack
Hobb es is a teacher than Bill
Comparative adjectives such as better come from the lexicon as compar By having trees
such as An transmit the compar value of the A no de to the ro ot N no de we can signal to
CnxPnx that it may adjoin when a comparative adjective has adjoined An example of such
an adjunction is given in Figure Of course if no comparative element is present in the
lower part of the noun phrase nxPnx will not be able to adjoin since nouns themselves are
compar In order to capture the fact that a comparativeelement blo cks further mo di cation
to N An must only adjoin to N no des whichare compar in their lower feature matrix
NP r
NP f PP NA
Nr P NP
A Nf than N NA
better teacher Bill
Figure Adjunction of nxPnx to NP mo di ed by comparative adjective
In order to obtain this result for phrases like more exquisite horse we need to provide a
way for more and less to mo dify adjectives without a than clause as we have with ARBaPa
Actually we need this ability indep endently for comparative adjectival phrases as discussed in
the next section
Adjectival Comparatives
With nominal comparatives we saw that a single analysis was amenable to b oth pure com
paratives and elliptical comparatives This is not p ossible for adjectival comparatives as the
following examples demonstrate
The dog is less patient
The dog is less patient than the cat
t The dog is as patien
CHAPTER COMPARATIVES
The dog is as patientasthecat
The less patient dog waited eagerly for its master
The less patient than the cat dog waited eagerly for its master
The last example shows that comparative adjectival phrases cannot distribute quite as freely
as comparative nominals
The analysis of elliptical comparative adjectives follows closely to that of comparative nom
inals We build them up by rst adjoining the comparative element to the A no de which then
signals to the AP no de via the compar feature that it may allowa than clause to adjoin The
relevant trees are given in Figure CARBa is anchored by more less and as and axPnx
is anchored byboththan and as
APr
Ar APf* PP
Ad◊ Af*
NA P◊ NP ↓
a CARBa tree b axPnx tree
Figure Elliptical adjectival comparative trees
The advantages of this analysis are many We capture the distribution exhibited in the
examples given in With CARBa comparative elements may mo dify adjectives
wherever they o ccur However than clauses for adjectives have a more restricted distribution
which coincides nicely with the distribution of AP s in the XTAG grammar Thus by making
them adjoin to AP rather than A ill formed sentences like are not allowed
There are two further advantages to this analysis One is that CARBa interacts with
nxPnx to pro duce sequences lik e more exquisite horse than Black Beauty a result alluded
to at the end of Section We achieve this by ensuring that the comparativeness of an
adjective is controlled by a comparative adverb which adjoins to it A sample derivation is
given in Figure The second advantage is that we get sentences suchas for free
Hobb es is b etter than Bill
Since better comes from the lexicon as compar and this value is passed up to the AP no de
axPnx can adjoin as desired giving us the derivation given in Figure
Notice that the ro ot AP no de of Figure is compar so we are basically saying that
e This accords with our use of the compar strings suchasbetter than Bil l are not comparativ
feature a p ositive value for compar signals that the clause b eneath it is to be compared NP r
NP f PP
Nr P NP
Ar Nf than Nr NA
Ad Af horse A Nf NA NA
more exquisite Black Beauty
Figure Comparativized adjective triggering CnxPnx
against something else In the case of better than Bil l the comparison has b een ful lled so we
do not want it to signal for further comparisons A nice result which follows is that axPnx
cannot adjoin more than once to anygiven AP spine and wehave no need for the NA constraint
on the tree s ro ot no de Also this treatment of the comparativeness of various strings proves
imp ortant in getting the co ordination of comparative constructions to work prop erly
A note needs to be made ab out the analysis regarding the interaction of the equivalence
comparative construction as as and the inequivalence comparative construction more less
than In the grammar more less and as all anchor CARBa and both than and as
anchor axPnx Without further mo di cations this of course will give us sentences such as the
following
Hobb es is as patient than Bill
Hobb es is more patient as Bill
Such cases are blo cked with the feature equiv more less fewer and than are equiv while
as in b oth adverbial and prep ositional uses is equiv The prep ositional trees then require
that their P no de and the no de to which they are adjoining matchforequiv
An interesting phenomena in which comparisons seem to be paired with an inappropriate
as than clause is exhibited in and
Hobb es is as patient or more patient than Bill
CHAPTER COMPARATIVES
APr compar : - NA
APf wh : <1> - PP NA compar : <2> +
A compar : <2> P NP wh : <1>
better than N
Bill
Figure Adjunction of axPnx to comparative adjective
Hobb es is more patient or as patient as Bill
Though prescriptive grammars disfavor these sentences these are p erfectly acceptable Wecan
capture the fact that the as than clause shares the equiv value with the latter of the comparison
phrases by passing the equiv value for the second element to the ro ot of the co ordination tree
Adverbial Comparatives
The analysis of adverbial comparatives encouragingly parallels the analysis for nominal and
elliptical adjectival comparatives with however some interesting di erences Some examples
of adverbial comparatives and their distribution are given in the following
Alb ert works more quickly
Alb ert works more quickly than Richard
Alb ert works more
Alb ert more works
Alb ert works more than Richard
Hobb es eats his supp er more quickly than Calvin
Hobb es more quickly eats his supp er than Calvin
Hobb es more quickly than Calvin eats his supp er
When more is used alone as an adverb it must also occur after the verb phrase Also it
app ears that adverbs mo di ed by more and less have the same distribution as when they are
not mo di ed However the than p ortion of an adverbial comparative is restricted to p ost verb
phrase p ositions
The rst observation can b e captured byhaving more and less select only vxARB from the
set of adverb trees Comparativization of adverbs lo oks very similar to that of other categories
and we follow this trend by giving the tree in Figure a which parallels the adjectival and
nominal trees for these instances This handles the quite free distribution of adverbs which
have b een comparativized while the tree in Figure b vxPnx allo ws the than p ortion of
an adverbial comparative to o ccur only after the verb phrase blo cking examples such as
VPr
Adr VPf* PP
Ad◊ Adf*
NA P◊ NP ↓
a CARBarb tree b vxPnx tree
Figure Adverbial comparative trees
The usage of the compar feature parallels that of the adjectives and nominals however
trees which adjoin to VP are compar on their ro ot VP no de In this way vxPnx anchored
by than or as which must adjoin to a compar VP can only adjoin immediately ab ove
a comparative or comparativized adverb This avoids extra parses in which the comparative
adverb adjoins at a VP no de lower than the than clause
A nal note is that as may anchor vxPnx non comparativ ely as in sentence This
means that there will b e two parses for sentences suchas
John works as a carp enter
John works as quickly as a carp enter
This app ears to be a legitimate ambiguity One is that John works as quickly as a carp enter
works quickly and the other is that John works quickly when he is acting as a carp enter but
mayb e he is slow when he acting as a plumber
CHAPTER COMPARATIVES
Future Work
Interaction with determiner sequencing e g several more men than women but not every
more men than women
Handle sentential complement comparisons e g Bil l eats more pasta than Angus drinks
beer
Add partitives
Deal with constructions like as many and as much
Lo ok at so as construction
Chapter
Punctuation Marks
Many parsers require that punctuation b e stripp ed out of the input Since punctuation is often
optional this sometimes has no e ect However there are a number of constructions which
must obligatorily contain punctuation and adding analyses of these to the grammar without
the punctuation would lead to severe overgeneration An esp ecially common example is noun
app ositives Without access to punctuation one would have to allow every combinatorial
p ossibility of NPs in noun sequences which is clearly undesirable esp ecially since there is
already unavoidable noun noun comp ounding ambiguity Aside from coverage issues it is
also preferable to take input as is and do as little editing as p ossible With the addition
of punctuation to the XTAG grammar we need only do assume the conversion of certain
sequences of punctuation into the British order this is discussed in more detail below in
Section
The XTAG POS tagger currently tags every punctuation mark as itself These tags are all
converted to the POS tag Punct b efore parsing This allows us to treat the punctuation marks
as a single POS class They then have features which distinguish amongst them Wherever
p ossible wehave the punctuation marks as anchors to facilitate early ltering
The full set of punctuation marks is separated into three classes balanced separating
and terminal The balanced punctuation marks are quotes and parentheses separating are
commas dashes semi colons and colons and terminal are p erio ds exclamation p oints and
question marks Thus the punct feature is complex like the agr feature yielding
feature equations like Punct bal paren or Punct term excl Separating and
terminal punctuation marks do not o ccur adjacent to other memb ers of the same class but may
o ccasionally occur adjacent to members of the other class e g a question mark on a clause
Balanced punctuation marks are sometimes which is separated by a dash from a second clause
adjacent to one another e g quotes immediately inside of parentheses The punct feature
allows us to control these lo cal interactions
We also need to control non lo cal interaction of punctuation marks Two cases of this
are so called quote alternation wherein emb edded quotation marks must alternate between
single and double and the imp ossibility of emb edding an item containing a colon inside of
another item containing a colon Thus we have a fourth value for punct contains
colon dquote etc which indicates whether or not a constituentcontains a particular
is p ercolated through all auxiliary trees Things which may punctuation mark This feature
not embed are colons under colons semi colons dashes or commas semi colons under semi
CHAPTER PUNCTUATION MARKS
colon or commas Although it is rare parentheses may app ear inside of parentheses saywith
a bibliographic reference inside a parenthesized sentence
App ositives parentheticals and vo catives
These trees handle constructions where additional lexical material is only licensed in conjunction
with particular punctuation marks Since the lexical material is unconstrained virtually any
noun can o ccur as an app ositive the punctuation marks are anchors and the other no des are
substitution sites There are cases where the lexical material is restricted as with parenthetical
adverbs like however and in those cases wehave the adverb as the anchor and the punctuation
marks as substitution sites
When these constructions can app ear inside of clauses non p eripherally they must b e sep
arated by punctuation marks on b oth sides However when they o ccur p eripherally they have
either a preceding or following punctuation mark We handle this by having both p eripheral
and non p eripheral trees for the relevant constructions The alternative is to insert the second
following punctuation mark in the tokenization pro cess i e insert a comma b efore the p erio d
when an app ositive app ears on the last NP of a sentence However this is very di cult to do
accurately
nxPUnxPU
The symmetric non p eripheral tree for NP app ositives anchored by comma dash or paren
theses It is shown in Figure anchored by parentheses
The music here Russell Smith s Tetrameron sounded go o d Brown cc
cost million p ounds million dollars
Sen David Boren D Okla
some analysts b elieve the two recent natural disasters Hurricane Hugo and the San
Francisco earthquake will carry economic rami cations WSJ
The punctuation marks are the anchors and the app ositive NP is substituted The app ositive
can be conjoined but only with a lexical conjunction not with a comma App ositives with
commas or dashes cannot b e pronouns although they may b e conjuncts containing pronouns
When used with parentheses this tree actually presents an alternative rather than an app ositive
so a pronoun is p ossible Finally the app ositive p osition is restricted to having nominativeor
accusative case to blo ckPRO from app earing here
App ositives can b e emb edded as in but do not seem to b e able to stack on a single
NP In this they are more like restrictive relatives than app ositive relatives whichtypically can
stack
noted Simon Brisco e UK economist for Midland Montagu a unit of Midland Bank
PLC NPr
NPf* Punct1 NPr Punct2 NA
( D NPf ) NA
3 Nr
N Nf NA
million dollars
Figure The nxPUnxPU tree anchored by parentheses
nPUnxPU
The symmetric non p eripheral tree for N level NP app ositives is anchored by comma The
mo di er is typically an address It is clear from examples such as that these are attached
at N rather than NP Carrier is not an app ositive on Menlo Park as it would be if these
were simply stacked app ositives Rather Calif mo di es Menlo Park and that entire complex
is comp ounded with carrier as shown in the correct derivation in Figure Because this
distinction is less clear when the mo di er is p eripheral e g ends the sentence and it would b e
di cult to distinguish b etween NP and N attachment we do not currently allowa p eripheral
N level attachment
An o cial at Consolidated Freightways Inc a Menlo Park Calif less than truckload
carrier said
Rep Ronnie Flipp o D Ala of the delegation says
nxPUnx
This tree which can b e anchored by a comma dash or colon handles asymmetric p eripheral
NP app ositives and NP colon expansions of NPs Figure shows this tree anchored by a
dash and a colon Like the symmetric app ositive tree nxPUnxpu the asymmetric app ositive
cannot be a pronoun while the colon expansion can Thus this constraint comes from the
syntactic entry in b oth cases rather than b eing built into the tree
the bank s shareholder Petroliam Nasional Bhd Brown
said Chris Dillow senior U K economist at Nomura Research Institute
CHAPTER PUNCTUATION MARKS
NPr
NPf Punct1 NPr Punct2 NA
N , D NPf , NA
Consolidated a Nr
Nr Nf NA
Nf Punct1 NP Punct2 carrier NA
Menlo-Park , N ,
Calif
Figure An N level mo di er using the nPUnx tree
qualities that are seldom found in one work Scrupulous scholarship a fund of p ersonal
exp erience Brown cc
I had eyes for only one p erson him
The colon expansion cannot itself contain a colon so the fo ot S has the feature NP t
punctcontainscol on
PUpxPUvx
Tree for pre VP parenthetical PP anchored by commas or dashes
John in a t of anger brokethevase
Mary just within the last year has totalled twocars
These are clearly not NP mo di ers
Figures and show this tree alone and as part of the parse for
puARBpuvx
Parenthetical adverbs however though etc Since the class of adverbs is highly restricted this
tree is anchored by the adverb and the punctuation marks substitute The punctuation marks Sr
NPr VPr
D NPf Ad VP NA NA
our N always V NPr NPr
family HATED NPf Punct NP
D NPr NA
NPr G NPf Punct NP N : Nr NA
cats A Nf D NP ’s N - N f r r NA NA
A1 Conj A things the N N Nf N Nf NA NA NA
A1 Conj A , vulgar bank 90% shareholder Paetroliam N Nf NA NA
Naasional Bhd nasty , low
a b
Figure The derived trees for an NP with a a p eripheral dash separated app ositiveand
b an NP colon expansion uttered by the Mouse in Alice s Adventures in Wonderland
VPr
Punct1 PP↓ Punct2 VP* NA
, ,
Figure The PUpxPUvx tree anchored bycommas
may be either commas or dashes Like the parenthetical PP ab ove these are clearly not NP
mo di ers
CHAPTER PUNCTUATION MARKS
Sr
NP VPr
N Punct1 PP Punct2 VP NA
John , P NPr , V NPr
in D NPf broke D NPf NA NA
a NPf PP the N NA
N P NP vase
fit of N
anger
Figure Tree illustrating the use of PUpxPUvx
The new argumentover the noti cation guideline however could sour any atmosphere
of co op eration that existed WSJ
sPUnx
Sentence nal vocative anchored by comma
You were there Stanley myboy
Also when anchored by colon NP expansion on S These often app ear to be extrap osed
mo di ers of some internal NP The NP must b e quite heavy and is usually a list
Of the ma jor expansions in three were nanced under the R I Industrial Building
Authority s mortgage plan Collyer Wire Leesona Corp oration and American Tub e
Controls
A simpli ed version of this sentence is shown in gure The NP cannot b e a pronoun
in either of these cases Both vo catives and colon expansions are restricted to app ear on tensed
clauses indicative or imp erative Sr
S Punct NP NA
NPr VPr : NP1 Conj NP NA
D NPf V VP NP1 Conj NP and Nr NA NA NA
three N were VP PP Nr , Nr N Nf NA NA
expansions V P NPr N Nf N Nf American Controls NA NA
financed under D NPf Collyer Wire Leesona Corporation NA
the N
plan
Figure A tree illustrating the use of sPUnx for a colon expansion attached at S
nxPUs
Tree for sentence initial vo catives anchored by a comma
Stanley myboy you were there
The noun phrase may be anything but a pronoun although it is most commonly a prop er
noun The clause adjoined to must b e indicative or imp erative
Bracketing punctuation
Simple bracketing
Trees PUsPU PUnxPU PUnPU PUvxPU PUvPU PUarbPU PUaPU PUdPU
PUpxPU PUpPU
These trees are selected by parentheses and quotes and can adjoin onto anynodetyp e whether
a head or a phrasal constituent This handles things in parentheses or quotes which are syn
tactically integrated into the surrounding context Figure shows the PUsPU anchored by
parentheses and this tree along with PUnxPU in a derived tree
Dick Carroll and his accordion which we now refer to as Freida held over at Bahia
Cabana where Sir Judson Smith brings in his calypso cap ers Oct Brown ca
CHAPTER PUNCTUATION MARKS
noted that the term teacher employee as opp osed to e g maintenance employee
was a not inapt description Brown ca
NPr
NPf Sr NA
D NPf Punct1 Sf Punct2 NA NA
his N ( Comp Sr ) NA
accordion which NP VPr
Sr N VP PP NA
we V PP1 P NPr
refer P NP as Punct NP Punct Punct1 Sf* Punct2 1 1 f 2 NA NA NA
to ε ‘‘ N ’’
( ) Freida
a b
Figure PUsPU anchored by parentheses and in a derivation along with PUnxPU
There is a convention in English that quotes emb edded in quotes alternate between single
and double in American English the outermost are double quotes while in British English they
are single The contains feature is used to control this alternation The trees anchored by
double quotation marks have the feature punct contains dquote on the fo ot no de and
the feature punct containsdquote on the ro ot All adjunction trees are transparentto
the contains feature so if any tree b elow the double quote is itself enclosed in double quotes
the derivation will fail Likewise with the trees anchored by single quotes The quote trees in
ximity is handled by the punct e ect toggle the contains Xquote feature Immediate pro
balanced feature which allows quotes inside of parentheses but not vice versa
In addition American English typically places moves p erio ds and commas inside of quo
tation marks when they would logically o ccur outside as in example The comma in the
rst part of the quote is not part of the quote but rather part of the parenthetical quoting
clause However by convention it is shifted inside the quote as is the nal p erio d British
English do es not do this We assume here that the input has already b een tokenized into the
British format
You can t do this to us Diane screamed We are Americans
The PUsPU can handle quotation marks around multiple sentences since the sPUs tree
allows us to join two sentences with a p erio d exclamation point or question mark Currently
however we cannot handle the style where only an op en quote app ears at the b eginning of a
paragraph when the quotation extends over multiple paragraphs We could allowa lone op en
quote to select the PUs tree if this is deemed desirable
Also the PUsPU is selected by a pair of commas to handle non p eripheral app ositive
relative clauses such as in example Restrictive and app ositive relative clauses are not
syntactically di eren tiated in the XTAG grammar cf Chapter
This news announced by Jerome To obin the orchestra s administrative director
brought applause Brown cc
The trees discussed in this section will only allow balanced punctuation marks to adjoin to
constituents We will not get them around non constituents as in
Mary asked him to leave and he left
sPUsPU
This tree allows a parenthesized clause to adjoin onto a non parenthesized clause
Innumerable motels from Tucson to New York boast swimming p o ols swim at your
own risk is the hospitable sign p oised at the brink of most p o ols Brown ca
Punctuation trees containing no lexical material
PU
This is the elementary tree for substitution of punctuation marks This tree is used in the
quoted sp eech trees where including the punctuation mark as an anchor along with the verb
of saying would require a new entry for every tree selecting the relevant tree families It is also
used in the tree for parenthetical adverbs puARBpuvx and for S adjoined PPs and adverbs
spuARB and spuPnx
PUs
Anchored by comma allows comma separated clause initial adjuncts
Here as in Journal Mr Louis has given himself the lion s share of the dancing
Brown cc
Choreographed byMr Nagrin the work lled the second half of a program
Tokeep this tree from app earing on ro ot Ss i e sentence wehave a ro ot constraint that
punct struct nil similar to the requirement that ro ot Ss b e tensed i e mo de ind imp
The punct struct nilfeature on the fo ot blo cks stacking of multiple punctuation marks
This feature is shown in the tree in Figure
This tree can b e also used by adjuncts on emb edded clauses
CHAPTER PUNCTUATION MARKS
Sr invlink : <1> inv : <1> punct : struct : <2> displ-const : <3> agr : <4> assign-case : <5> mode : <6> sub-conj : <7> extracted : <8> tense : <9> assign-comp : <10> comp : <11>
Punct punct : struct : <2> S* inv : <1> punct : punct : struct : comma NA struct : nil displ-const : <3> agr : <4> assign-case : <5> mode : <6> sub-conj : <7> extracted : <8> tense : <9> assign-comp : <10> comp : <11>
,
Figure PUs with features displayed
One might exp ect that in a p o etic career of seventy o dd years some changes in style and
metho d would have o ccurred some developmenttaken place Brown cj
These adjuncts sometimes have commas on b oth sides of the adjunct or like only
have them at the end of the adjunct
Finally this tree is also used for p eripheral app ositive relative clauses
Interest may remain limited into tomorrow s U K trade gures which the market will
be watching closely to see if there is anyimprovement after disapp ointing numb ers in the
previous twomonths
sPUs
This tree handles clausal co ordination with comma dash colon semi colon or any of the
terminal punctuation marks The rst clause must be either indicative or imp erative The
second may also b e in nitival with the separating punctuation marks but must b e indicativeor
imp erative with the terminal marks with a comma it may only b e indicative The two clauses
need not share the same mo de NB Allowing the terminal punctuation marks to anchor this
tree allows us to parse sequences of multiple sentences This is not the usual mo de of parsing
if it were this sort of sequencing might b e b etter handled by a higher level of pro cessing
For critics Hardy has had no p o etic p erio ds one do es not sp eak of early Hardy or late
Hardy or of the London or Max Gate p erio d
Then there was exercise b oating and hiking whichwas not only go o d for you but also
made you more virile the thoughtofstrenuous activity left him exhausted
This construction is one of the few where two non bracketing punctuation marks can be
adjacent It is p ossible if rare for the rst clause to end with a question mark or exclamation
p oint when the two clauses are conjoined with a semi colon colon or dash Features on the
fo ot no de as shown in Figure control this interaction
Complementizers are not p ermitted on either conjunct Sub ordinating conjunctions some
times app ear on the right conjunct but seem to b e imp ossible on the left
Killpath would just have to go out and drag Gun backby the heels once an hour b ecause
he d b e damned if he was going to b e a mid watch p encil pusher Brown cl
The best rule of thumb for detecting corked wine provided the eye has not already
sp otted it is to smell the wet end of the cork after pulling it if it smells of wine
if it smells of cork one has grounds for suspicion the bottle is probably all right
Brown cf
sPU
This tree handles the sentence nal punctuation marks when selected by a question mark
exclamation p oint or period One could also require a nal punctuation mark for all clauses
but such an approachwould not allow non p erio ds to o ccur internally for instance b efore a semi
colon or dash as noted ab ove in Section This tree currently only adjoins to indicative
or imp erative ro ot clauses
He left
Get lost
Get lost
The feature punct bal nil on the fo ot no de ensures that this tree only adjoins inside
theses or quotes completely enclosing a sentence but do es not restrict it from of paren
adjoining to clause which ends with balanced punctuation if only the end of the clause is
contained in the parentheses or quotes
CHAPTER PUNCTUATION MARKS
Sr wh : <1> displ-const : <2> agr : <3> assign-case : <4> mode : <5> ind/imp tense : <6> assign-comp : <7> comp : <8> nil punct : contains : <9>
↓ Sf* mode : <5> Punct punct : contains : <9> S1 comp : nil NA comp : <8> punct : struct : none punct : contains : scolon : + assign-comp : <7> contains : colon : - tense : <6> struct : scolon mode : ind/imp/inf assign-case : <4> agr : <3> displ-const : <2> wh : <1> sub-conj : nil punct : struct : none term : excl/qmark contains : colon : -
;
Figure sPUs with features displayed
John then left
John then left
Mary asked him to leave immediately
This tree is also selected by the colon to handle a colon expansion after adjunct clause
Expressed di erently if the price for b ecoming a faithful follower Brown cd
Expressing it di erently if the price for b ecoming a faithful follower
To express it di erently if the price for b ecoming a faithful follower Brown cd
This tree is only used after adjunct untensed clauses which adjoin to the tensed clause
using the adjunct clause trees cf Section the mo de of the complete clause is that of the
matrix rather than the adjunct Indicative or imp erative i e ro ot clauses separated by a
colon use the sPUs tree Section
vPU
This tree is anchored by a colon or a dash and o ccurs between a verb and its complement
These typically are lists
Printed material Available on request from U S Department of Agriculture Wash
ington D C are Co op erativeFarm Credit Can Assist Brown ch
pPU
This tree is anchored by a colon or a dash and o ccurs b etween a prep osition and its complement
It typically o ccurs with a sequence of complements As with the tree ab ove this typically o ccurs
with a conjoined complement
and utilization suchas A the protection of forage
can b e represented as Af
Other trees
spuARB
In general we attach p ost clausal mo di ers at the VP no de as you typically get scop e ambiguity
e ects with negation John didn t leave to day did he leave or not However with p ost
sentential comma separated adverbs there is no ambiguity in John didn t leave today he
de nitely did not leave Since this tree is only selected by a subset of the adverbs namely
those which can app ear pre sententially without a punctuation mark it is anchored by the
adverb
The names of some of these pro ducts don t suggest the risk involved in buying them
either WSJ
spuPnx
Clause nal PP separated by a comma Like the adverbs describ ed ab ove these di er from VP
adjoined PPs in taking widest scop e
gold for current delivery settled at an ounce up cents
It increases employee commitment to the company with all that means for e ciency
and qualitycontrol
nxPUa
Anchored by colon or dash allows for p ost mo di cation of NPs by adjectives
Make no mistake this Gorky Studio drama is a resp ectable imp ort aptly grave
carefully written p erformed and directed
CHAPTER PUNCTUATION MARKS
Part V
App endices
App endix A
Future Work
A Adjective ordering
At this point the treatment of adjectives in the XTAG English grammar do es not include
selectional or ordering restrictions Consequently any adjective can adjoin onto any noun and
on top of any other adjective already mo difying a noun All of the mo di ed noun phrases shown
in currently parse
big green bugs
big green ideas
colorless green ideas
green big ideas
While are all semantically anomalous also su ers from an ordering prob
lem that makes it seem ungrammatical as well Since the XTAG grammar fo cuses on syntactic
constructions it should accept but not Both the auxiliary and determiner
ordering systems are structured on the idea that certain typ es of lexical items sp eci ed by fea
tures can adjoin onto some typ es of lexical items but not others We b elieve that an analysis
of adjectival ordering would follow the same typ e of mechanism
A More work on Determiners
In addition to the analysis describ ed in Chapter there remains work to b e done to complete
the analysis of determiner constructions in English Although constructions such as determiner
co ordination are easily handled if overgeneration is allowed blo cking sequences suchas one and
some while allowing sequences such as ve or ten still remains to be worked out There are
still a handful of determiners that are not currently handled by our system We do not have
This section is a rep eat of information found in section
This section is from Ho ckey and Mateyak
APPENDIX A FUTURE WORK
an analysis to handle most such certain other and own In addition there is a set of lexical
items that we consider adjectives enough less more and much that have the prop erty that
they cannot co o ccur with determiners We feel that a complete analysis of determiners should
b e able to account for this phenomenon as well
A ing adjectives
An analysis has already b een provided for past participal ed adjectives as in sentence
which are restricted to the TransitiveVerb family A similar analysis needs to take place for
the present participle ing used as a pre nominal mo di er This typ e of adjective however
do es not seem to be as restricted as the ed adjectives since verbs in other tree families seem
to exhibit this alternation as well e g sentences and
The murdered man was a do ctoral studentatUPenn
The man died
The dying man pleaded for his life
A Verb selectional restrictions
Although we explicitly do not want to mo del semantics in the XTAG grammar there is some
work along the syntax semantics interface that would help reduce syntactic ambiguity and
thus decrease the number of semantically anomalous parses In particular verb selectional
restrictions particularly for PP arguments and adjuncts would be quite useful With the
exception of the required to in the Ditransitive with PP Shift tree family Tnx Vnx Pnx
any prep osition is allowed in the tree families that have prep ositions as their arguments In
addition there are no restrictions as to which prep ositions are allowed to adjoin onto a given
verb The sentences in are all currently accepted by the XTAG grammar Their
violations are stronger than would b e exp ected from purely semantic violations however and
sentences from b eing the presence of verb selectional restrictions on PP s would keep these
accepted
survivors walked of the street
The man ab out the earthquake survived
The president arranged on a meeting
The behavior of own is su ciently unlike other determiners that it most likely needs a tree of its own
adjoining onto the right hand side of genitive determiners
This analysis may need to b e extended to the TransitiveVerb particle family as well
A Thematic Roles
Elementary trees in TAGs capture several notions of lo cality with the most primary of these
b eing lo cality of role assignment Each elementary tree has asso ciated with it the roles
assigned by the anchor of that elementary tree In the currentXTAG system while the notion
of lo calityof role assignment within an elementary tree has b een implicit the roles assigned
by a head have not been explicitly represented in the elementary tree Incorp orating role
information will make the elementary trees more informative and will enable e cient pruning
of spurious derivations when embedded into a sp eci c context In the case of a Synchronous
TAG roles can also be used to automatically establish links between two elementary trees
one in the ob ject language and one in the target language
App endix B
Metarules
B Intro duction
XTAG has now a collection of functions accessible from the user interface that helps the user in
the construction and maintenance of a tag tree grammar This subsystem is based on the idea of
metarules Becker Here our primary purp ose is to describ e the facilities implemented
under this metarule based subsystem For a discussion of the metarules as a metho d for compact
representation of the Lexicon see Becker and Srinivas et al
The basic idea of using metarules is to take pro t of the similarities of the relations involving
related pairs of XTAG elementary trees For example in the English grammar describ ed in
this technical rep ort comparing the XTAG trees for the basic form and the wh sub ject moved
form the relation between this two trees for transitive verbs nx Vnx W nx Vnx is
similar to the relation for the intransitiveverbs nx V W nx V and also to the relation for
the ditransitives nx Vnx nx W nx Vnx nx Hence instead of generating by hand the
six trees mentioned ab ove a more natural and robust way would be generating by hand only
the basic trees for the intransitive transitive and ditransitive cases and letting the wh sub ject
moved trees to b e automatically generated bytheapplication of a unique transformation rule
that would account exactly for the identical relation involved in each of the three pairs ab ove
Notice that the degree of generalization can be much higher than it might be thought in
principle from the ab ove paragraph For example once a rule for passivization is applied to the
tree di erent basic trees ab ove the wh sub ject mo ved rule could b e again applied to generate
the wh moved sub ject versions for the passive form Dep ending on the degree of regularity that
one can nd in the grammar b eing built the reduction in the numb er of original trees can be
exp onential
We still make here a pointthat the reduction of e ort in grammar construction is not the
only advantage of the approach Robustness reliability and maintainability of the grammar
achieved by the use of metarules are equally or even more imp ortant
In the next section we de ne a metarule in XTAG Section gives some linguistically
motivated examples of metarule for the English grammar describ ed in this technical rep ort and
their application Section describ es the access through the user interface
B The de nition of a metarule in XTAG
A metarule sp eci es a rule for transforming grammar rules into grammar rules In XTAG the
grammar rules are lexicalized trees Hence an XTAG metarule mr is a pair lhs rhs of XTAG
trees where
lhs the left hand side of the metarule is a pattern tree i e it is intended to present a
sp eci c pattern of tree to lo ok for in the trees submitted to the application of the metarule
When a metarule mr is applied to an input tree inp the rst step is to verify if the input
tree matches the pattern sp eci ed bythelhs If there is no match the application fails
rhs the right hand side of the metarule sp eci es together with lhs the transformation
that will b e done in inp in case of successful matching thus generating the output tree
of the metarule application
B No de names variable instantiation and matches
We will use the terms lhs rhs and inp asintro duced ab ove to refer to the parts of a generic
metarule b eing applied to an input tree
The no des at lhs can take three di erent forms a constant no de a typ ed variable no de
and anon typ ed variable no de The naming conventions for these di erent classes of no des is
given b elow
Constant No de Its name must not initiate bya question mark character They
are likewe exp ect for names to b e in normal XTAG trees for instance inp is exp ected
to have only constant no des Some examples of constant no des are NP V NP NP
S We will call the two parts that comp ose such names the stem and the subscript In
r
the examples ab ove NP V and S are stems and r are subscripts Notice that the
subscript part can also b e emptyasintwo of the ab ove examples
Non Typ ed Variable No de Its name initiates by a question mark followed bya
sequence of digits i e a numb er which uniquely identi es the variable Examples
We assume that there is no stem and no subscript in this names i e is
just a meta character to intro duce a variable and the number is the v ariable identi er
Typ ed Variable No de Its name initiates by a question mark followed by a
sequence of digits but is additionally followed by a type speci ers de nition A type
speci ers de nition is a sequence of one or more type speci er separated by a slash
A typespeci er has the same form of a regular XTAG no de name like the constantnodes
except that the subscript can b e also a question mark Examples of typ ed variables are
VP a single typ e sp eci er with stem VP and no subscript NP P P two typ e
sp eci ers NP and PP NP one typ e sp eci er NP with undetermined subscript
actually more than one output tree can b e generated from the successful application of a rule to an input
tree as will b e seen so on
Notice however that having the sole purp ose of distinguishing b etween variables a number like the one in
the last example is not very likely to o ccur and a metarule with more than three thousand variables can give
you a place in the Guinness TagBo ok of Records
APPENDIX B METARULES
We ll see ahead that each typ e sp eci er represents an alternative for matching and the
presence of in subscript p osition of a typ e sp eci er means that matching will only
check for the stem
During the pro cess of matching variables are asso ciated we use the term instantiated with
tree material According to its class a variable can b e instantiated with di erent kinds of tree
material
Atyp ed variable will b e instantiated with exactly one no de of the input tree whichisin
accordance to one of its typ e sp eci ers The full rule is in the following subsection
A non typ ed variable will b e instantiated with a range of subtrees These subtrees will
b e taken from one of the no des of the input tree inp Hence there will a no de n in inp
with subtrees n t n t n t in this order where the variable will b e instantiated with
k
some subsequence of these subtrees e g n t n t n t Note however that some of
these subtrees may b e incomplete i e they may not go all the way to the b ottom leaves
Entire subtrees may b e removed Actually for eachchild of the non typ ed variable no de
one subtree that matches this child subtree will b e removed from some of the n t mayb e
i
an entire n t leaving in place a mark for inserting material during the substitution of
i
o ccurences at rhs
Notice still that the variable can b e instantiated with a single tree and even with no tree
We de ne a match to b e a complete instantiation of all variables app earing in the metarule
In the pro cess of matching there may be several p ossible ways of instantiating the set of
variables of the metarule i e several p ossible matches This is due to the presence of non
typ ed variables
Now we are ready to de ne what we mean by a successful matching The pro cess of
matching is successful if the number of p ossible matches is greater then When there is no
p ossible match the pro cess is said to fail In addition to return success or failure the pro cess
also return the set of all p ossible matches which will b e used for generating the output
B Structural Matching
The pro cess of matching lhs and inp can b e seen as a recursive pro cedure for matching trees
starting at their ro ots and pro ceeding in a top down style along with their subtrees In the
explanation of this pro cess that follows we have used the term lhs not only to refer to the
whole tree that contains the pattern but to any of its subtrees that is b eing considered in a
given recursive step The same applies to inp By nowwe ignore feature equations which will
be accounted for in the next subsection
The pro cess describ ed b elow returns at the end the set of matches where an empty set
means the same as failure We rst give one auxiliary de nition of valid Mapping and one
recursive function Match that matches lists of trees instead of trees and then de ne the pro cess
of matching two trees as a sp ecial case of call to Match
This is di erent from not having a subscript whichisinterpreted as checking that the input tree haveno
subscript for matching
Given a list list lhs lhs l hs ofnodesof lhs and a list l ist inp inp inp
inp i
lhs l
of no des of inp we de ne a mapping from l ist to list to b e a function M apping that for
inp
lhs
each elementof list assigns a list of elements of l ist de ned by the following condition
inp
lhs
concatenation M apping lhs M apping lhs M apping lhs list
inp
l
i e the elements of list are split into sublists and assigned in order of app earance in the
inp
list to the elements of list
lhs
We say that a mapping is a valid mapping if for all j j l where l is the length of
list the following restrictions apply
lhs
if lhs is a constant no de then M apping lhs must have a single element say rhs
j j
g j
and the two no des must have the same name and agree on the markers fo ot substitution
head and NA i e if lhs is NA then rhs must b e NA if lhs has no markers then
j j
g j
rhs must havenomarkers etc
g j
if lhs is a typ e variable no de then M apping lhs must have a single element say
j j
rhs and rhs must b e marker compatible and type compatible with lhs
j
g j g j
rhs is marker compatible with lhs if any marker fo ot substitution head and NA
j
g j
presentinlhs is also presentinrhs
j
g j
rhs is type compatible with lhs if there is at least one of the alternativetyp e sp eci ers
j
g j
for the typ ed variable that satis es the conditions b elow
has the stem de ned in the typ e sp eci er rhs
g j
if the typ e sp eci er do esn t have subscript then rhs must have no subscript
g j
if the typ e sp eci er has a subscript di erent of then rhs must have the same
g j
subscript as in the typ e sp eci er
if lhs is a non typ ed variable no de then there s actually no requirement M apping lhs
j j
mayhave any length and even b e empty
The following algorithm Match takes as input a list of no des of lhs and a list of no des of
inp and returns the set of p ossible matches generated in the attempt of match this two lists
If the result is an empty set this means that the matching failed
Function Match list l ist
lhs rhs
Let MAPP INGS b e the list of all valid mappings from l ist to list
lhs rhs
Make M ATCHES
For each mapping M apping MAPP INGS do
Make M atches f g
For each j j l where l l eng th l ist do
lhs
if lhs is a constant no de then
j
Notice that unlike the case for the constant no de the inverse is not required i e if lhs has no marker
j
rhs is still allowed to have some
g j
If the typ e sp eci er has a subscript there is no restriction and that is exactly its function to allow for
the matching to b e indep endent of the subscript
APPENDIX B METARULES
let chil dr en b e the list of children of lhs
j
lhs
lhr b e the single elementinM apping lhs
j
g j
chil dr en b e the list of children of lhr
rhs
g j
Make Matches fm m j m Matches
j
and m Match chil dr en chil dr en g
j
lhs rhs
if lhs is a typ ed variable no de then
j
let chil dr en b e the list of children of lhs
j
lhs
lhr b e the single elementinM apping lhs
j
g j
chil dr en b e the list of children of lhr
rhs
g j
Make Matches ff lhs lhr g m m j m Matches
j j
g j
and m Match chil dr en chil dr en g
j
lhs rhs
if lhs is a non typ ed variable no de then
j
let chil dr en b e the list of children of lhs
j
lhs
sl b e the number of nodes in chil dr en
lhs
DESC b e the set of s size lists given by
s
DE S C f dr dr dr j
s s
for every k s dr is a descendant
k
of some no de in M apping lhs
j
for every k s dr is to the rigth of dr
l k
ery list Desc dr dr dr DE S C do For ev
s s
Let Tree Material b e the list of subtrees dominated
by the no des in M apping lhs but with the
j
subtrees dominated by the no des in DE S C
s
cut out from these trees
Make Matches ff lhs Tree Struct g m m j
j j
m M atches and m Match chil dr en D esc g
j
lhs
Make MATCHES MATCHES M atches
Return M ATCHES
Finally we can de ne the pro cess of structurally matching lhs to inp as the evaluation of
Match ro ot lhs ro ot inp If the result is an empty set the matching failed otherwise the
resulting set is the set of p ossible matches that will b e used for generating the new trees after
b eing pruned by the feature equation matc hing
B Output Generation
Although nothing has yet b een said ab out the feature equations which is the sub ject of the
next subsection we assume that only matches that meet the additional constraints imp osed by
feature equations are considered for output If no structural match survives feature equations
checking that matching has failed
If the pro cess of matching lhs to inp fails there are two alternativebehaviors according to
the value of a parameter If the parameter is set to false whichisthe default value no output
is generated On the other hand if it is set to true then the own inp tree is copied to the
output
the parameter is accessible at the Lisp interface by the name XTAG metarules copy unmatched trees
If the pro cess of matching succeeds as many trees will b e generated in the output as the
number of p ossible matches obtained in the pro cess For a given match the output tree is
generated by substituting in the rhs tree of the metarule the o ccurrences of variables by the
material to which they have b een instantiated in the match The case of the typ ed variable
is simple The name of the variable is just substituted by the name of the no de to which it
hasbeeninstantiated from inp Avery imp ortant detail is that the marker fo ot substitution
head NA or none at the output tree no de comes from what is sp eci ed in the rhs no de which
can b e di erent of the marker at the variable no de in inp and of the asso ciated no de from inp
The case of the non typ ed variable not surpringly is not so simple In the output tree this
no de will b e substituted by the subtree list that was asso ciated to this no de in the same other
attaching to the parentofthisnon typ ed variable no de But rememb er that some subtrees may
have been removed from some of the trees in this list mayb e entire elements of this list due
to the e ect of the children of the metavariable in lhs It is a requirementthatany o ccurence
of a non typ ed variable no de at the rhs tree has exactly the same number of children than the
unique o ccurence of this non typ ed variable no de in lhs Hence when generating the output
tree the subtrees at rhs will be inserted exactly at the p oints where subtrees were removed
during matching in a p ositional one to one corresp ondance
For feature equations in the output trees see the next subsection The comments at the
output are the comments at the lhs tree of the metarule followed by the coments at inp both
parts intro duced by appropriate headers allowing the user to have a complete history of each
tree
B Feature Matching
In the previous subsections we have considered only the asp ects of a metarule involving the
structural part of the XTAG trees In a feature based grammar as XTAG is accounting for
tial A metarule is not really worth if it do esn t account for the prop er change features is essen
of feature equations from the input to the output tree The asp ects that have to b e considered
here are
Which feature equations should b e required to b e presentininp in order for the match
to succeed
Which feature equations should b e generated in the output tree as a function of the feature
equations in the input tree
Based on the p ossible combinations of these requirements we partition the feature equations
into the following ve classes
Require Retain Feature equations in this class are required to be in inp in order for
matching to succeed Up on matching these equations will b e copied to the output tree
Notice that what is really imp ortant is not the features themselves but the feature equations that relate the
feature values of no des of the same tree
This classi cation is really a partition i e no equation may b e conceptually in more than one class at the
same time
APPENDIX B METARULES
To achieve this b ehaviour the equation must be placed in the lhs tree of the metarule
preceded by a plus character e g V t tr ans
Require Don t Copy The equation is required to be in inp for matching but should
not be copied to the output tree Those equations must be in lhs preceded by minus
character e g NP case acc
Optional Don t Copy The equation is not required for matching but wehave to make
sure not to copy it to the output tree set of equations regardless of it b eing present or
not in inp Those equations must b e in lhs in raw form i e neither preceded byaplus
nor minus character e g S b per f ect VP t per f ect
r
Optional Retain The equation is not required for matching but in case it is found
in inp it must be copied to the output tree This is the default case and hence these
equations should not b e present in the metarule sp eci cation
Add The equation is not required for matching but we want it to be put in the output
tree anyway These equations are placed in rawforminthe rhs notice in this case it is
the right hand side
Typ ed variables can b e used in feature equations in b oth lhs and rhs They are intended to
represent the no des of the input tree to which they have b een instantiated For each resulting
match from the structural matching pro cess the following is done
The typ ed variables in the equations at lhs and rhs are substituted by the names of
the no des they have b een instantiated to
The requirements concerning feature equations are checked according to the ab ove rules
If the match survives feature equation checking the prop er output tree is generated
according to Section B and to the rules describ ed ab ove for the feature equations
be intro duced in Finally a new kind of metavariable which is not used at the no des can
the feature equations part They have the same form of the non typ ed variables i e quotation
mark followed by a numb er and are used in the place of feature values and feature names
Hence if the equation NP b app ears in lhs this means that all feature equations
of inp that match a b ottom attribute of some NP to any feature value but not to a feature
path will not b e copied to the output
B Examples
Figure B shows a metarule for wh movement of the sub ject Among the trees to whichithave
been applied are the basic trees of intransitive transitive and ditransitive families including
prep ositional complements passive trees of the same families and ergative
Commutativity of equations is accounted for in the system Hence an equation x y can also b e sp eci ed
as y x Asso ciativity is not accounted for and its need by an user is viewed as indicating missp eci cation at
the input trees
lhs rhs
Figure B Metarule for wh movement of sub ject
Figure B shows a metarule for wh movement of an NP in ob ject p osition Among the
trees to whichithave b een applied are the basic and passive trees of transitive and ditransitive
families
lhs rhs
Figure B Metarule for wh movementofobject
Figure B shows a metarule for general wh movement of an NP It can be applied to
APPENDIX B METARULES
generate trees with either sub ject or ob ject NP moved We show in Figure B the basic tree
for the family Tnx Vnx Pnx and the tree wh trees generated by the application of the rule
lhs rhs
Figure B Metarule for general wh movement of an NP
B The Access to the Metarules through the XTAG Interface
We rst describ e the access to the metarules subsystem using bu ers with single metarule
applications Then we pro ceed by describing the application of multiple metarules in what we
call the parallel sequential and cumulative mo des to input tree les
Wehave de ned conceptually a metarule as an ordered pair of trees In the implementation
of the metarule subsystem it works the same a metaruleisabu erwithtwo trees The name
of the metarule is the name of the bu er The rst tree that app ear in the main window under
the metarule bu er is the left hand side the next app earing b elow is the right hand side The
p ositional approach allows us to have naming freedom the tree names are irrelevant Since
wecansavebu ersinto text les we can talk also ab out metarule les
The available options for applying a metarule which is in a bu er are
For applying it to a single input tree clickinthename of the tree in the main window
and cho ose the option apply metarule to tree You will be prompted for the name of
the metarule to apply to the tree which should b e as wementioned b efore the name of
the bu er that contains the metarule trees The output trees will be generated at the
end of the bu er that contains the input tree The names of the trees dep end of a LISP
parameter metarules change name If the value of the parameter is false the default
Although a bu er is intended to implement the concept of a set not a sequence of trees wetake pro t of
the actual organization of the system to realize the concept of ordered tree pair in the implementation
so that even if wewanttohave mnemonic names resembling their distinct character left or right hand side
wehave some naming exibility to call them e g lhs or lhs passive
Tnx Vnx Pnx sub ject moved
NP ob ject moved NP ob ject moved from PP
Figure B Application of wh movement rule to Tnx Vnx Pnx
value then the new trees will have the same name as the input otherwise the name of
the input tree followed by a dash and the name of the right hand side of the tree
The value of the parameter can b e changed bycho osing Tools at the menu bar and then
the reason whywe do not use the name of the metarule i e the name of the bu er is b ecause in some forms
of application the metarules do not carry individual names as we ll see so on is the case when a set of metarules
from a le is applied
APPENDIX B METARULES
either name mr output trees input or append rhs name to mr output trees
For applying it to all the trees of a bu er click in the name of the bu er that contains
the trees and pro ceed as ab ove The output will b e a new bu er with all the output trees
The name of the new bu er will b e the same as the input bu er pre xed by MR The
names of the trees follow the conventions ab ove
The other options concern application to les instead of bu ers Lets rst de ne the
concepts of parallel sequential and cumulative application of metarules One metarule le can
contain more than one metarule The rst two trees i e the rst tree pair form one metarule
lets call it mr Subsequent pairs in the sequence of trees de ne additional metarules mr
mr mr
n
We say that a metarule le is applied in parallel to a tree see Figure B if eachofthe
metarules is applied indep endently to the input generating its particular output trees
We generalize the concept to the application in parallel of a metarule le to a tree le
with p ossibly more than one tree generating all the trees as if each metarule in the
metarule le was applied to each tree in the input le
Input Trees
MR0 MR1 MRn
Output
Trees
Figure B Parallel application of metarules
We say that a metarule le mr mr mr mr is applied in sequence to a input tree
n
le see Figure B if we apply mr to the trees of the input le and for each i n
apply metarule mr to the trees generated as a result of the application of mr
i i
Input MR0 MR1 MRn Output
Trees Trees
Figure B Sequential application of metarules
Finally the cumulative application is similar to the sequential except that the input trees
at each stage are by passed to the output together with the newly generated ones see
Figure B
rememb er a metarule application generates as many output trees as the numb er of matches Input MR0 MR1 MRn Output
Trees Trees
Figure B Cumulative application of metarules
Remember that in case of matching failure the output result is decided as explained in
subsection B either to be empty or to be the input tree The re ex here of having the
parameter set for copying the input is that for the parallel application the output will haveas
many copies of the input as matching failures For the sequential case the decision apply at
each level and setting the parameter for copying in a certain sense guarantees for the pip e
not to break Due to its nature and unlike the two other mo des the cumulative application is
not a ected by this parameter
The options for application of metarules to les are available by clicking at the menu item
Tools and then cho osing the appropriate function among
Apply metarule to les You ll be prompted for the metarule le name which should
contain one metarule and for input le names Each input le name inp le will be
indep endently submitted to the application of the metarule generating an output le with
the name MR inp le
Apply metarules in paral lel to les You ll b e prompted for the metarules le name with
one or more metarules and for input le names Each input le name inp le will be
indep endently submitted to the application of the metarules in parallel For each parallel
application to a le inp le an output le with the name MRP inp le will b e generated
Apply metarules in sequence to les The interaction is as describ ed for the application in
parallel except that the application of the metarules are in sequence and that the output
les are pre xed by MRS instead of MRP
Apply metarules cumulatively to les The interaction is as describ ed for the applications
parallel and in sequence except that the mo de of application is cumulative and that in
the output les are pre xed by MRC
Finally still under the Tools menuwe can change the setting of the parameter that controls
the output result on matching failure see Subsection B by cho osing either copy input on
mr matching failure or no output on mr matching failure
if it contains more than trees the additional trees are ignored
App endix C
Lexical Organization
C Intro duction
An imp ortant characteristic of an FB LTAG is that it is lexicalized i e each lexical item is
anchored to a tree structure that enco des sub categorization information Trees with the same
canonical sub categorizations are group ed into tree families The reuse of tree substructures
such as wh movement in many di erent trees creates redundancy which p oses a problem for
grammar development and maintenance Vijay Shanker and Schab es To consistently
implement a change in some general asp ect of the design of the grammar all the relevant
trees currently must b e insp ected and edited Vijay Shanker and Schab es suggested the use of
hierarchical organization and of tree descriptions to sp ecify substructures that would b e present
in several elementary trees of a grammar Since then in addition to ourselves Becker Becker
Evans et al Evans et al and Candito Candito havedevelop ed systems for
organizing trees of a TAG which could b e used for developing and maintaining grammars
Our system is based on the ideas expressed in Vijay Shanker and Schab es Vijay Shanker
and Schab es to use partial tree descriptions in sp ecifying a grammar by separately
de ning pieces of tree structures to enco de indep endentsyntactic principles Various individual
sp eci cations are then combined to form the elementary trees of the grammar The chapter
its implementation We begins with a description of our grammar development system and
will then show the main results of using this to ol to generate the Penn English grammar as
well as a Chinese TAG We describ e the signi cant prop erties of b oth grammars p ointing out
the ma jor di erences b etween them and the metho ds bywhich our system is informed ab out
these language sp eci c prop erties The chapter ends with the conclusion and future work
C System Overview
In our approach three typ es of comp onents sub categorization frames blo cks and lexical
redistribution rules are used to describ e lexical and syntactic information Actual trees
are generated automatically from these abstract descriptions as shown in Figure C In
ever be manipulated the tree maintaining the grammar only the abstract descriptions need
descriptions and the actual trees which they subsume are computed deterministically from
these high level descriptions Subcategorization frames
Lexical redistribution rules(LRRs) Our System Sets of Trees Subcategorization blocks
Transformation
blocks
Figure C Lexical Organization System Overview
C Sub categorization frames
Sub categorization frames sp ecify the category of the main anchor the number of arguments
each argument s category and p osition with resp ect to the anchor and other information such
as feature equations or no de expansions Each tree family has one canonical sub categorization
frame
C Blo cks
Blo cks are used to represent the tree substructures that are reused in di erent trees i e blo cks
subsume classes of trees Each blo ck includes a set of no des dominance relation parent relation
precedence relation b etween no des and feature equations This follows the de nition of the tree
descriptions sp eci ed in a logical language patterned after Rogers and Vijay Shanker Rogers and
Vijay Shankar
Blo cks are divided into two typ es according to their functions sub categorization blo cks
and transformation blo cks The former describ es structural con gurations incorp orating the
various information in a sub categorization frame For example some of the sub categorization
blo cks used in the development of the English grammar are shown in Figure C
When the sub categorization frame for a v erb is given by the grammar develop er the system
will automatically create a new blo ck of co de byessentially selecting the appropriate primi
tive sub categorization blo cks corresp onding to the argument information sp eci ed in that verb
frame
The transformation blo cks are used for various transformations such as wh movement
These transformation blo cks do not enco de rules for mo difying trees but rather describ e the
prop erties of a particular syntactic construction Figure C depicts our representation of
phrasal extraction This can be sp ecialized to give the blo cks for wh movement topicaliza
tion relative clause formation etc For example the wh movement blo ck is de ned by further
In order to fo cus on the use of tree descriptions and to make the gures less cumb ersome we show only
the structural asp ects and do not show the feature value sp eci cation The parent immediate dominance
relationship is illustrated by a plain line and the dominance relationship by a dotted line The arc b etween no des
shows the precedence order of the no des are unsp eci ed The no des categories are enclosed in parentheses
APPENDIX C LEXICAL ORGANIZATION sp ecifying that the ExtractionRo ot is lab eled S the NewSite has a wh feature and so on
Root Root VP(’VP’) Subject(’NP’) PredP Subject VP Verb(’V’) Pred (a) is_main_frame (b)pred_has_subject (c) subject_is_NP
PredP VP,PredP Object(’NP’) Pred Object V, Pred
(f)object_is_NP
(d) main_pred_is_V (e)pred_has_object
Figure C Some sub categorization blo cks
ExtRoot
NewSite Root(’S’)
ExtNode
ε
(a) extraction
NPr(’NP’)
NPf(’NP’) SB(’S’) ExtRoot(’S’) ExtRoot(’S’) NewSite Root(’S’) COMP(’COMP’)
ε NewSite Root(’S’) ExtNode ExtNode ε ε
(b) wh-movement (c) relative clause
Figure C Transformation blo cks for extraction
C Lexical Redistribution Rules LRRs
The third typ e of machinery available for a grammar develop er is the Lexical Redistribution
Rule LRR An LRR is a pair r r of sub categorization frames which pro duces a new frame
r
l
when applied to a sub categorization frame s by rst matching the left frame r of r to s then
l
combining information in r and s LRRs are intro duced to incorp orate the connection b etween
r
sub categorization frames For example most transitiveverbs have a frame for active a sub ject
and an ob ject and another frame for passive where the ob ject in the former frame b ecomes
the sub ject in the latter An LRR denoted as passive LRR is built to pro duce the passive
sub categorization frame from the active one Similarly applying dative shift LRR to the frame
with one NP sub ject and two NP ob jects will pro duce a frame with an NP sub ject and an PP
ob ject
Besides the distinct content LRRs and blo cks also di er in several asp ects
They have di erent functionalities Blo cks represent the substructures that are reused in
di erent trees They are used to reduce the redundancy among trees LRRs are intro duced
to incorp orate the connections b etween the closely related sub categorization frames
Blo cks are strictly additive and can be added in any order LRRs on the other hand
pro duce di erent results dep ending on the order they are applied in and are allowed to
be non additive i e to remove information from the sub categorization frame they are
b eing applied to as in the pro cedure of passive from active
Sq S Sq q
NP S NP NP S S
NP VP 0 NP VP NP VP 0 1
VNP1 PP2 V VNP12 PP PP2 ε P 22NP P 22NP P 22NP
to to ε to ε
(a) (b) (c)
Figure C Elementary trees generated from combining blo cks
C Tree generation
To generate elementary trees we b egin with a canonical sub categorization frame The system
will rst generate related sub categorization frames by applying LRRs then select sub cate
gorization blo cks corresp onding to the information in the sub categorization frames next the
combinations of these blo cks are further combined with the blo cks corresp onding to various
Matching occurs successfully when frame s is compatible with r in the typ e of anchors the number of
l
arguments their p ositions categories and features In other words incompatible features etc will blo ck certain
LRRs from b eing applied
APPENDIX C LEXICAL ORGANIZATION
transformations nally a set of trees are generated from those combined blo cks and they are
the tree family for this sub categorization frame Figure C shows some of the trees pro duced
in this way For instance the last tree is obtained by incorp orating information from the di
transitiveverb sub categorization frame applying the dative shift and passive LRRs and then
combining them with the wh non sub ject extraction blo ck Besides in our system the hierarchy
for sub categorization frames is implicit as shown in Figure C
subject_is_NP object_is_NP main_anchor_is_verb object_is_AP
Intransitive(walk) Intrans. with adj (feel happy) Transitive(buy) Intrans. particle(add up)
Ditrans. with NP Trans. particle(pick up) (ask) Ditrans. with PP (put) Trans. idioms Ditrans. with S (kick the bucket) . (force sb to do sth) . Trans. light verb . (take a walk)
.
Figure C Partial inheritance lattice in English
C Implementation
The input of our system is the description of the language which includes the sub categorization
frame list LRR list sub categorization blo ck list and transformation lists The output is a list
of trees generated automatically by the system as shown in Figure C The tree generation
mo dule is written in Prolog and the rest part is in C Wealsohave a graphic interface to input
the language description Figure C and C are two snapshots of the interface
C Generating grammars
We have used our to ol to sp ecify a grammar for English in order to pro duce the trees used in
the current English XTAG grammar Wehave also used our to ol to generate a large grammar
for Chinese In designing these grammars wehave tried to sp ecify the grammars to re ect the
similarities and the di erences b etween the languages The ma jor features of our sp eci cation
of these two grammars are summarized in Table C and C
Both grammars are still under development so the contents of these two tables mightchange a lot in the
future according to the analyses we cho ose for certain phenomenon For example the ma jority of work on
Chinese grammar treat ba construction as some kind of ob ject fronting where the character ba is either an ob ject
marker or a prep osition According to this analysis an LRR rule for ba construction is used in our grammar
to generate the preverbal ob ject frame from the p ostverbal frame However there has b een some argument for
treating ba as a verb If we later cho ose that analysis the main verbs in the patterns NP VP and NP ba
NP VP will b e di erent therefore no LRR will b e needed for it As a result the numb ers of LRRs sub cat
frames and tree generated will change accordingly Lexorg system subsets of derived frames subcat blocks Apply Select Combine w/ trans LRRs subcat blocks blocks
combinations of subsets of subcat blocks and trans. blocks
Tree generation
subcat frame list trees generated LRR list subcat block list elementary trees trans block list for the language
Language description
Figure C Implementation of the system
Figure C Interface for creating a grammar
By fo cusing on the sp eci cation of individual grammatical information we have b een able
to generate nearly all of the trees from the tree families used in the current English grammar
APPENDIX C LEXICAL ORGANIZATION
Figure C Part of the Interface for creating blo cks
English Chinese
examples passive b ei construction
of LRRs dative shift ob ject fronting
ergative ba construction
examples wh question topicalization
of transformation relativization relativization
blo cks declarative argument drop
LRRs
sub cat blo cks
trans blo cks
sub cat frames
trees generated
Table C Ma jor features of English and Chinese grammars
develop ed at Penn Our approach has also exp osed certain gaps in the Penn grammar We
are encouraged with the utility of our to ol and the ease with which this large scale grammar
was develop ed
We are currently working on expanding the contents of sub categorization frame to include
trees for other categories of words For example a frame which has no sp eci er and one NP
complement and whose predicate is a prep osition will corresp ond to PP P NP tree We ll
also intro duce a mo di er eld and semantic features so that the head features will propagate
Wehavenotyet attempted to extend our coverage to include punctuation it clefts and a few idiosyncratic
analyses
b oth grammars English Chinese
causative long passive VO inversion
LRRs short passive ergative ba const
dative shift
topicalization
trans blo cks relativization gerund argument drop
declarative
NP S sub ject zero sub ject
sub cat blo cks S NP PP ob ject PL ob ject preverbal ob ject
V predicate prep predicate
Table C Comparison of the two grammars
from mo di ee to mo di ed no de while non head features from the predicate as the head of the
mo di er will b e passed to the mo di ed no de
C Summary
Wehave describ ed a to ol for grammar development in which tree descriptions are used to pro
vide an abstract sp eci cation of the linguistic phenomena relevant to a particular language
In grammar development and maintenance only the abstract sp eci cations need to b e edited
and anychanges or corrections will automatically b e proliferated throughout the grammar In
addition to lightening the more tedious asp ects of grammar maintenance this approach also
allows a unique p ersp ectiveon the general characteristics of a language De ning hierarchical
blo cks for the grammar both necessitates and facilitates an examination of the linguistic as
sumptions that have b een made with regard to feature sp eci cation and tree family de nition
This can be very useful for gaining an overview of the theory that is being implemented and
exp osing gaps that remain unmotivated and need to be investigated The typ e of gaps that
can b e exp osed could include a missing sub categorization frame that might arise from the au
tomatic combination of blo cks and whichwould corresp ond to an entire tree family a missing
tree whichwould represent a particular typ e of transformation for a sub categorization frame
or inconsistent feature equations By fo cusing on syntactic prop erties at a higher level our
approach allows new opp ortunities for the investigation of how languages relate to themselves
and to each other
App endix D
Tree Naming conventions
The various trees within the XTAG grammar are named more or less according to the following
tree naming conventions Although these naming conventions are generally followed there are
o ccasional trees that do not strictly follow these conventions
D Tree Families
Tree families are named according to the basic declarative tree structure in the tree family see
section D but with a T as the rst character instead of an or
D Trees within tree families
Each tree b egins with either an alpha or a b eta symb ol indicating whether it is an
initial or auxiliary tree resp ectively Following an or a the name may additionally contain
one of
I imp erative
E ergative
N relative clausefp ositiong
G NP gerund
D Determiner gerund
pW wh PP extraction fp ositiong
W wh NP extractionfpositiong
X ECM eXceptional case marking
Numbers are assigned according to the p osition of the argument in the declarative tree as
follows
sub ject p osition
rst argument e g direct ob ject
second argument e g indirect ob ject
The b o dy of the name consists of a string of the following comp onents which corresp onds to
the leaves of the tree The anchor s of the trees is are indicated by capitalizing the part of
sp eech corresp onding to the anchor
s sentence
a adjective
arb adverb
be be
c relative complementizer
x phrasal category
d determiner
v verb
lv lightverb
conj conjunction
comp complementizer
it it
n noun
p prep osition
to to
pl particle
by by
neg negation
As an example the transitive declarative tree consists of a sub ject NP followed byaverb which
is the anchor followed by the ob ject NP This translates into nx Vnx If the sub ject NP
had b een extracted then the tree would b e W nx Vnx A passive tree with the by phrase in
the same tree family would b e nx Vbynx Note that even though the ob ject NP has moved
to the sub ject p osition it retains the ob ject enco ding nx
D Assorted Initial Trees
Trees that are not part of the tree families are generally gathered into several les for conve
nience The various initial trees are lo cated in lex trees All the trees in this le should b egin
with an indicating that they are initial trees This is followed by the ro ot category which
follows the naming conventions in the previous section e g n for noun x for phrasal category
The ro ot category is in all capital letters After the ro ot category the no de leaves are named
b eginning from the left with the anchor of the tree also b eing capitalized As an example the
NXN tree is ro oted by an NP no de NX and anchored by a noun N
D Assorted Auxiliary Trees
The auxiliary trees are mostly lo cated in the bu ers prepositions trees conjunctions trees
determiners trees advs adjs trees and modifiers trees although a couple of other les
also contain auxiliary trees The auxiliary trees follow a sligh tly di erent naming convention
from the initial trees Since the ro ot and fo ot no des must b e the same for the auxiliary trees
the ro ot no des are not explicitly mentioned in the names of auxiliary trees The trees are
named according to the leaf no des starting from the left and capitalizing the anchor no de All
auxiliary trees b egin with a of course For example ARBs indicates a tree anchored by
APPENDIX D TREE NAMING CONVENTIONS
an adverb ARB that adjoins onto the left of an S no de Note that S must b e the fo ot no de
and therefore also the ro ot no de
D Relative Clause Trees
For relative clause trees the following naming conventions have been adopted if the wh
moved NP is overt it is not explicitly represented Instead the index of the site of movement
for sub ject for ob ject for indirect ob ject is app ended to the N So N nx Vnx
is a sub ject extraction relative clause with NP substitution and N nx Vnx is an ob ject
w
extraction relative clause If the wh moved NP is covert and Comp substitutes in the Comp
no de is represented by c in the tree name and the index of the extraction site follows c Thus
Nc nx Vnx is a sub ject extraction relative clause with Comp substitution Adjunct trees
are similar except that since the extracted material is not co indexed to a trace no index
is sp eci ed cf Npxnx Vnx which is an adjunct relative clause with PP pied piping and
Ncnx Vnx which is an adjunct relative clause with Comp substitution Cases of pied
piping in which the pied pip ed material is part of the anchor ha ve the anchor capitalized or
sp elled out cf Nbynx nx Vbynx whichisa relative clause with by phrase pied piping and
NP substitution
w
App endix E
Features
Table E contains a comprehensive list of the features in the XTAG grammar and their p ossible
values
This section consists of short biographical sketches of the various features currently in use
in the XTAG English grammar
E Agreement
hagri is a complex feature It can have as its subfeatures
hagr rdsingi p ossible values
hagr numi p ossible values pl ur sing
hagr persi p ossible values
hagr geni p ossible values masc f em neut
These features are used to ensure agreement between a verb and its sub ject
Where do es it o ccur
Nouns comes sp eci ed from the lexicon with their hagri features e g books is hagr rdsingi
hagr numi plur andhagr persi Only pronouns use the gen gender feature
The hagri features of a noun are transmitted up the NP tree by the following equation
NP b hagri N t hagri
Agreementbetween a verb and its sub ject is mediated by the following feature equations
NP hagri VP t hagri
subj
VP b hagri V t hagri
Agreement has to b e done as a two step pro cess b ecause whether the verb agrees with the
sub ject or not dep ends up on whether some auxiliary verb adjoins in and up on what the hagri
sp eci cation of the verb is
Verbs also come sp eci ed from the lexicon with their hagri features e g the hagri features
of the verb sings are hagr rdsingi hagr numi sing andhagr p ersi Non nite forms
of the verb sing e g singing do not come with an hagri feature sp eci cation
APPENDIX E FEATURES
Feature Value
agr rdsing
agr num plur sing
agr p ers
agr gen fem masc neuter
assign case nom acc none
assign comp that whether if for ecm rel inf nil ind nil ppart nil none
card
case nom acc gen none
comp that whether if for rel inf nil ind nil nil
compar
compl
conditional
conj and or but comma scolon to disc nil
const
contr
control no value indexing only
decreas
de nite
displ const
equiv
extracted
gen
gerund
inv
invlink no value indexing only
irrealis
mainv
mo de base ger ind inf imp nom ppart prep sb junt
neg
passive
p erfect
pred
progressive
pron
punct bal dquote squote paren nil
punct contains colon
punct contains dash
punct contains dquote
punct contains scolon
punct contains squote
punct struct comma dash colon scolon nil
punct term p er qmark excl nil
quan
re
rel clause
rel pron ppart ger adj clause
select mo de ind inf ppart ger
sup er
tense pres past
trace no value indexing only
trans
weak
wh
Table E List of features and their p ossible values
E Agreement and Movement
The hagri features of a moved NP and its trace are co indexed This captures the fact that
movement do es not disrupt a pre existing agreement relationship b etweenanNPandaverb
Whichboys do es John think t are is intelligent
i i
E Case
There are two features resp onsible for case assignment
hcasei p ossible values nom acc gen none
hassign casei possiblevalues nom acc none
Case assigners prep ositions and verbs as well as the VP S and PP no des that dominate
them haveanhassign casei case feature Phrases and lexical items that havecasei e Ns and
i feature NPs haveahcase
Case assignmentby prep ositions involves the following equations
PP b hassign casei P t hcasei
NP t hcasei P t hcasei
Prep ositions come sp eci ed from the lexicon with their hassign casei feature
P b hassign casei acc
Case assignmentbyverbs has two parts assignment of case to the ob ject s and assignment
of case to the sub ject Assignment of case to the ob ject is simpler English verbs always assign
accusative case to their NP ob jects direct or indirect Hence this is built into the tree and
not put into the lexical entry of each individual verb
NP t hcasei acc
obj ect
Assignment of case to the sub ject involves the follo wing two equations
NP hcasei VP t hassign casei
subj
VP b hassign casei V t hassign casei
This is a two step pro cess the nal case assigned to the sub ject dep ends up on the hassign
casei feature of the verb as well as whether an auxiliary verb adjoins in
Finite verbs like sings have nom as the value of their hassign casei feature Non nite
verbs have none as the value of their hassign casei feature So if no auxiliary adjoins in the
only sub ject they can have is PRO which is the only NP with none as the value its hcasei
feature
APPENDIX E FEATURES
E ECM
Certain verbs e g want believe consider etc and one complementizer for are able to assign
case to the sub ject of their complement clause
The complementizer for like the prep osition for has the hassign casei feature of its com
plement set to acc Since the hassign casei feature of the ro ot S of the complement tree and
r
the hcasei feature of its NP sub ject are co indexed this leads to the sub ject b eing assigned
accusative case
ECM verbs have the hassign casei feature of their fo ot S no de set to acc The co indexation
between the hassign casei feature of the ro ot S and the hcasei feature of the NP sub ject leads
r
to the sub ject b eing assigned accusative case
E Agreement and Case
The hcasei features of a moved NP and its trace are co indexed This captures the fact that
movement do es not disrupt a pre existing relationship of case assignment between a verb and
an NP
Her She I think that Odo liket
i i i
E Extraction and Inversion
hextractedi p ossible vales are
All sentential trees with extracted comp onents with the exception of relative clauses are
marked S bhextractedi at their top S no de The extracted element may be a wh NP or
a topicalized NP Thehextractedi feature is currently used to blo ckemb edded topicalizations
as exempli ed by the following example
John w ants Bill PRO to leavet
i i
htracei this feature is not assigned any value and is used to co index moved NPs and their
traces which are marked by
hwhi p ossible values are
NPs like who what etc come marked from the lexicon with a value of for the feature
hwhi NPs hwhi Non wh NPs have as the value of their hwhi feature Note that
are not restricted to o ccurring in extracted p ositions to allow for the correct treatmentofecho
questions
The hwhi feature is propagated up by p ossessives e g the hwhi feature of the determiner
which in which boy is propagated up to the level of the NP so that the value of the hwhi feature
of the entire NP is hwhi This pro cess is recursive e g which boy s mother which boy s
mother s sister
The hwhi feature is also propagated up PPs Thus the PP to whom has as the value of
its hwhi feature
In trees with extracted NPs the hwhi feature of the ro ot no de S no de is equated with the
hwhi feature of the extracted NPs
constraints Certain verbs like The hwhi feature is used to imp ose sub categorizational
wonder can only take interrogative complements other verbs such as know can take both
interrogative and non interrogative complements and yet other verbs like think can only take
non interrogative complements cf the hextractedi and hmo dei features also play a role in
imp osing sub categorizational constraints
The hwhi feature is also used to get the correct inversion patterns
E Inversion Part
The following three features are used to ensure the correct pattern of inversion
hwhi p ossible values are
hinvi p ossible values are
hinvlinki p ossible values are
Facts to b e captured
No inversion with topicalization
No inversion with matrix extracted sub ject wh questions
Inversion with matrix extracted ob ject wh questions
Inversion with all matrix wh questions involving extraction from an emb edded clause
No inversion in emb edded questions
No matrix sub ject topicalizations
Consider a tree with ob ject extraction where NP is extracted The following feature equa
tions are used
S b hwhi NP t hwhi
q
S b hinvlinki S b hinvi
q q
S b hinvi S t hinvi
q r
S b hinvi
r
Ro ot restriction A restriction is imp osed on the nal ro ot no de of anyXTAGderivation of
the hinvlinki feature of the nal ro ot a tensed sentence which equates the hwhi feature and
no de
If the extracted NP is not a wh word i e its hwhi feature has the value at the end of
the derivation S b hwhi will also have the value Because of the ro ot constraint S b hwhi
q q
will b e equated to S b hinvlinki which will also come to have the value Then by
q
S t hinvi will acquire the value This will unify with S b h invi which has the value cf
r r
Consequently no auxiliary verb adjunction will be forced Hence there will never be
inversion in topicalization
If the extracted NP is a wh word i e its hwhi feature has the value at the end of the
derivation S b hwhi will also have the value Because of the ro ot constraint S b hwhi
q q
value Then by will b e equated to S b hinvlinki which will also come to have the
q
S t hinvi will acquire the value This will not unify with S b hinvi which has the value
r r
cf Consequently the adjunction of an inverted auxiliary verb is required for the
derivation to succeed
Inversion will still take place even if the extraction is from an emb edded clause
APPENDIX E FEATURES
Who do es Loida think Miguel likes t
i i
This is b ecause the adjoined tree s ro ot no de will also haveitsS b hinvi set to
r
Note that inversion is only forced up on us b ecause S is the nal ro ot no de and the Ro ot
q
restriction applies In embedded environments the ro ot restriction would not apply and the
feature clash that forces adjunction would not take place
The hinvlinki feature is not present in sub ject extractions Consequently there is no inver
sion in sub ject questions
Sub ject topicalizations are blo cked by setting the hwhi feature of the extracted NP to
i e only wh phrases can go in this lo cation
E Inversion Part
hdispl consti
Possible values set set
In the previous section we saw how inversion is triggered using the hinvlinki hinvi hwhi
features Inversion involves movement of the verb from V to C This movement pro cess is
represented using the hdispl consti feature which is used to simulate Multi Comp onentTAGs
The sub value set indicates the inversion multi comp onentset while there are not currently
any other uses of this mechanism it could b e expanded with other sets receiving di erent set
values
The hdispl consti feature is used to ensure adjunction of two trees which in this case
are the auxiliary tree corresp onding to the moved verb S adjunct and the auxiliary tree
corresp onding to the trace of the moved verb VP adjunct The following equations are used
S b hdispl const set i
r
S t hdispl const set i
VP b hdispl const set i V t hdispl const set i
V b hdispl const set i
S b hdispl const set i VP t hdispl const set i
r
E Clause Typ e
There are several features that mark clause typ e They are
hmo dei
hpassivei p ossible values are
i p ossible values are base ger ind inf imp nom ppart prep sb jnct hmo de
The hmo dei feature of a verb in its ro ot form is base The hmo dei feature of a verb in its
past participial form is ppart the hmo dei feature of a verb in its progressive gerundive form
The hdispl consti feature is also used in the ECM analysis
Wehave already seen one instance of a feature that marks clause typ e hextractedi which marks whether
a certain S involves extraction or not
is ger the hmo dei feature of a tensed verb is ind and the hmo dei feature of a verb in the
imp erativeis imp
nom is the hmo dei value of AP NP predicative trees headed by a null copula prep is the
hmo dei value of PP predicative trees headed by a null copula Only the copula auxiliary
tree some sentential complementverbs suchasconsider and raising verb auxiliary trees have
nom prep as the hmo dei feature sp eci cation of their fo ot no de This allow them and only
them to adjoin onto AP NP PP predicative trees with null copulas
E Auxiliary Selection
The hmo dei feature is also used to state the sub categorizational constraints between an aux
iliary verb and its complement We mo del the following constraints
have takes past participial complements
passive be takes past participial complements
active be takes progressive complements
mo dal verbs do and to take VPs headed byverbs in their base form as their complements
An auxiliary verb transmits its own mo de to its ro ot and imp oses its sub categorizational
restrictions on its complement i e on its fo ot no de e g the auxiliary have in its in nitival
form involves the following equations
VP b hmo dei V t hmo dei
r
V t hmo dei base
VP b hmo dei ppart
hpassivei This feature is used to ensure that passives only have be as their auxiliary Passive
trees start out with their hpassivei feature as This feature starts out at the level of the
verb and is p ercolated up to the level of the VP This ensures that only auxiliary verbs whose
fo ot no de has as their hpassivei feature can adjoin on a passive Passive trees have ppart
as the value of their hmo dei feature So the only auxiliary trees that we really have to worry
blo cking are trees whose fo ot no des have ppart as the value of their hmo dei feature ab out
There are two such trees the be tree and the have tree The be tree is ne b ecause its fo ot
no de has as its hpassivei feature so b oth the hpassivei and hmo dei values unify the have
tree is blo cked b ecause its fo ot no de has as its hpassivei feature
E Relative Clauses
Features that are p eculiar to the relative clause system are
hselect mo dei p ossible values are ind inf ppart ger
hrel proni p ossible values are ppart ger adj clause
hrel clausei p ossible values are
hselect mo dei
Comps are lexically sp eci ed for hselect mo dei In addition the hselect mo dei feature of a
Comp is equated to the hmo dei feature of its sister S no de by the following equation
APPENDIX E FEATURES
Comp t hselect mo dei S t hmo dei
t
The lexical sp eci cations of the Comps are shown b elow
Comp t hselect mo dei ind inf ger ppart
C
that Comp t hselect mo dei ind
for Comp t hselect mo dei inf
hrel proni
There are additional constraints on where the null Comp can o ccur The null Comp is not
C
p ermitted in cases of sub ject extraction unless there is an intervening clause or or the relative
clause is a reduced relative mo de ppart ger
To mo del this paradigm the feature hrel proni is used in conjunction with the following
equations
S t hrel proni Comp t hrel proni
r
S b hrel proni S b hmo dei
r r
Comp b hrel proni ppart ger adj clause for
C
The full set of the equations ab ove is only present in Comp substitution trees involving
sub ject extraction So the following will not b e ruled out
the toy Dafna likes t
i C i
The feature mismatch induced by the ab ove equations is not remedied by adjunction of just
any S adjunct b ecause all other S adjuncts are transparent to the hrel proni feature b ecause
of the following equation
S b hrel proni S t hrel proni
m
f
hrel clausei
The XTAG analysis forces the adjunction of the determiner b elowthe relative clause This is
done by using the hrel clausei feature The relevant equations are
On the ro ot of the RC NP b hrel clausei
r
On the fo ot no de of the Determiner tree NP t hrel clausei
f
E Complementizer Selection
The following features are used to ensure the appropriate distribution of complementizers
hcompi p ossible values that if whether for rel inf nil ind nil nil
hassign compi p ossible values that if whether for ecm rel ind nil inf nil none
hmo dei p ossible values ind inf sb jnct ger base ppart nom prep
hwhi p ossible values
The value of the hcompi feature tells us what complementizer we are dealing with The trees
which intro duce complementizers come sp eci ed from the lexicon with their hcompi feature
and hassign compi feature The hcompi of the Comp tree regulates what kind of tree go es
ab ove the Comp tree while the hassign compi feature regulates what kind of tree go es b elow
e g the following equations are used for that
S b hcompi Comp t hcompi
c
S b hwhi Comp t hwhi
c
S b hmo dei ind sb jnct
c
S t hassign compi Comp t hcompi
r
S b hcompi nil
r
By sp ecifying S b hcompi nil we ensure that complementizers do not adjoin onto other
r
complementizers The ro ot no de of a complementizer tree always has its hcompi feature set to
avalue other than nil
Trees that take clausal complements sp ecify with the hcompi feature on their fo ot no de
what kind of complementizer s they can take The hassign compi feature of an S no de is
tactic con guration the S no de is determined by the highest VP b elow the S no de and the syn
in
E Verbs with ob ject sentential complements
Finite sentential complements
S t hcompi that whether if nil
S t hmo dei ind sb jnct or S t hmo dei ind
nil inf nil S t hassign compi ind
The presence of an overt complementizer is optional
Non nite sentential complements do not p ermit for
S t hcompi nil
S t hmo dei inf
APPENDIX E FEATURES
S t hassign compi ind nil inf nil
Non nite sentential complements p ermit for
S t hcompi for nil
S t hmo dei inf
S t hassign compi ind nil inf nil
Cases like I want for to win are indep endently ruled out due to a case feature clash
between the acc assigned by for and the intrinsic case feature none on the PRO
Non nite sentential complements ECM
S t hcompi nil
S t hmo dei inf
S t hassign compi ecm
E Verbs with sentential sub jects
The following contrast involving complementizers surfaces with sentential sub jects
That John is crazy is likely
Indicative sentential sub jects obligatorily have complementizers while in nitival sentential
sub jects may or may not have a complementizer Also if is p ossible as the complementizer of
an ob ject clause but not as the complementizer of a sentential sub ject
S t hcompi that whether for nil
S t hmo dei inf ind
nil S t hassign compi inf
If the sentential sub ject is nite and a complementizer do es not adjoin in the hassign
compi feature of the S no de of the emb edding clause and the ro ot no de of the emb edded
clause will fail to unify If a complementizer adjoins in there will be no feature mismatch
b ecause the ro ot of the complementizer tree is not sp eci ed for the hassign compi feature
The hcompi feature nil is split into two hassign compi features ind nil and inf nil to
capture the fact that there are certain con gurations in which it is acceptable for an in nitival
clause to lack a complementizer but not acceptable for an indicative clause to lack a comple
mentizer
E That trace and for trace e ects
Who do you think that t ate the apple
i i
That trace violations are blo cked by the presence of the following equation
S b hassign compi inf nil ind nil ecm
r
nil feature on the b ottom of the S no des of trees with extracted sub jects W The ind
r
sp eci cation p ermits the ab ove example while the inf nil ecm feature sp eci cation allows the
following examples to b e derived
Who do you want t to win the World Cup
i i
Who do you consider t intelligent
i i
The feature equation that ruled out the that trace lter violations will also serve to rule out
the for trace violations ab ove
E Determiner ordering
hcardi p ossible values are
hcompli p ossible values are
hconsti p ossible values are
hdecreasi p ossible values are
hde nitei p ossible values are
hgeni p ossible values are
hquani p ossible values are
For detailed discussion see Chapter
E Punctuation
hpuncti is a complex feature It has the following as its subfeatures
hpunct bal i p ossible values are dquote squote paren nil
hpunct contains coloni p ossible values are
hpunct contains dashi p ossible values are
hpunct contains dquotei p ossible values are
hpunct contains scoloni p ossible values are
hpunct contains squotei p ossible values are
hpunct structi p ossible values are comma dash colon scolon none nil
hpunct termi p ossible values are per qmark excl none nil
For detailed discussion see Chapter
APPENDIX E FEATURES
E Conjunction
hconji p ossible values are but and or comma scolon to disc nil
The hconji feature is sp eci ed in the lexicon for each conjunction and is passed up to the ro ot
no de of the conjunction tree If the conjunction is and the ro ot hagr numi is hplurali no
matter what the numb er of the two conjuncts With or the the ro ot hagr numi is equated to
the hagr numi feature of the right conjunct
The hconji disc feature is only used at the ro ot of the CONJs tree It blo cks the
adjunction of one CONJs tree on another The follo wing equations are used where S is
r
the substitution no de and S is the ro ot no de
c
S t hconji disc
r
S b hconji and or but nil
c
E Comparatives
hcompari p ossible values are
hequivi p ossible values are
hsup eri p ossible values are
For detailed discussion see Chapter
E Control
hcontroli has no value and is used only for indexing purp oses Therootnodeofevery clausal
tree has its hcontroli feature coindexed with the control feature of its sub ject This allows
adjunct control to take place In addition clauses that take in nitival clausal complements
have the control feature of their sub ject ob ject coindexed with the control feature of their
complement clause S dep ending up on whether they are sub ject control verbs or ob ject control
verbs resp ectively
E Other Features
hnegi p ossible values are
Used for controlling the interaction of negation and auxiliary verbs
hpredi p ossible values are
The hpredi feature is used in the following tree families Tnx N trees and Tnx nx ARB trees
In the Tnx N trees family the following equations are used
for W nx N
NP t hpredi
NP b hpredi
NP t hpredi
N t hpredi NP b hpredi
This is the only tree in this tree family to use the hpredi feature
The other tree family where the hpredi feature is used is Tnx nx ARB trees Within this
family this feature and the following equations are used only in the W nx nx ARB tree
AdvP t hpredi
AdvP b hpredi
NP t hpredi
AdvP b hpredi NP t hpredi
hproni p ossible values are
This feature indicates whether a particular NP is a pronoun or not Certain constructions which
do not p ermit pronouns use this feature to blo ck pronouns
htensei p ossible values are pres past
It do es not seem to b e the case that the htensei feature interacts with other features syntactic
pro cesses It comes from the lexicon with the verb and is transmitted up the tree in such a
way that the ro ot Snode ends up with the tense feature of the highest verb in the tree The
equations used for this purp ose are
S b htensei VP t htensei
r
VP b htensei V t htensei
htransi p ossible values are
Many but not all English verbs can anchor b oth transitive and intransitive trees
The sun melted the ice cream
The ice cream melted
Elmo b orrowed a b o ok
A b o ok b orro wed
Transitive trees have the htransi feature of their anchor set to and intransitive trees
ve the htransi feature of their anchor set to Verbs such as melt which can o ccur in ha
b oth transitive and intransitive trees come unsp eci ed for the htransi feature from the lexicon
Verbs whichcan only o ccur in transitive trees e g borrow have their htransi feature sp eci ed
in the lexicon as thus blo cking their anchoring of an intransitivetree
App endix F
Evaluation and Results
In this app endix we describ e various evaluations done of the XTAG grammar Some of these
evaluations were done on an earlier version of the XTAG grammar the release while
other were done more recently We will try to indicate in each section whichversion was used
F Parsing Corp ora
In the XTAG pro ject we have used corpus analysis in two main ways to measure the
p erformance of the English grammar on a given genre and to identify gaps in the grammar
The second typ e of evaluation involves p erforming detailed error analysis on the sentences
rejected by the parser and wehave done this sev eral times on WSJ and Brown data Based on
the results of such analysis we prioritize up coming grammar development e orts The results
of a recent error analysis are shown in Table F The table do es not show errors in parsing
due to mistakes made by the POS tagger whichcontributed the largest number of errors
At this p oint wehave added a treatment of punctuation to handle an analysis of time NPs
a large number of multi word prep ositions part of gapless relative clauses
bare in nitives and have added the missing sub categorization and missing lexical
handle VP co ordination entry We are in the pro cess of extending the parser to
See Section on recentwork to handle VP and other predicative co ordination We nd that
this metho d of error analysis is very useful in fo cusing grammar development in a pro ductive
direction
To ensure that we are not losing coverage of certain phenomena as we extend the gram
mar wehave a b enchmark set of grammatical and ungrammatical sentences from this technical
rep ort We parse these sentences p erio dically to ensure that in adding new features and con
structions to the grammar we are not blo cking previous analyses There are approximately
example sentences in this set
F TSNLP
the English Grammar on the Test In addition to corpus based evaluation we have also run
Suites for Natural Language Pro cessing TSNLP English corpus Lehmann et al The
corpus is intended to b e a systematic collection of English grammatical phenomena including
Rank No of errors Category of error
Parentheticals and app ositives
Time NP
Missing sub cat
Multi word construction
Ellipsis
Not sentences
Relative clause with no gap
Funny co ordination
VP co ordination
Inverted predication
Who knows
Missing entry
Comparative
Bare in nitive
Table F Results of Corpus Based Error Analysis
complementation agreement mo di cation diathesis mo dality tense and asp ect sentence and
clause typ es co ordination and negation It contains grammatical sentences and phrases
and ungrammatical ones
Error Class Example
POS Tag She adds to V it He noises N him abroad
Missing lex item used as an auxiliary V calm NP down
Missing tree should ve bet NP NP S regard NP as Adj
Feature clashes My every rm All money
Rest approx e g
Table F Breakdown of TSNLP Errors
There were examples whichwe judged ungrammatical and removed from the test corpus
These were sentences with conjoined sub ject pronouns where one or both were accusative
e g Her and him succeed Overall we parsed of the remaining sentences and
phrases The errors were of various typ es broken down in Table F As with the error analysis
describ ed ab ove we used this information to help direct our grammar development e orts It
also highlighted the fact that our grammar is heavily slanted toward American English our
grammar did not handle dare or need as auxiliary verbs and there were a number of very British
particle constructions e g She misses him out
One general problem with the test suite is that it uses a very restricted lexicon and if
there is one problematic lexical item it is likely to app ear a large numb er of times and cause a
disprop ortionate amount of grief Used to app ears times and we got all wrong However
it must b e noted that the XTAG grammar has analyses for syntactic phenomena that were not
represented in the TSNLP test suite suchassentential sub jects and sub ordinating clauses among
others This e ort was therefore useful in highlighting some de ciencies in our grammar but
APPENDIX F EVALUATION AND RESULTS
did not provide the same sort of general evaluation as parsing corpus data
F Chunking and Dep endencies in XTAG Derivations
We evaluated the XTAG parser for the text chunking task Abney In particular we
compared NP chunks and verb group VG chunks pro duced by the XTAG parser with the
NP and VG chunks from the Penn Treebank Marcus et al The test involved
sentences of length words or less from sections to of the Penn Treebank parsed using
the XTAG English grammar The results are giv en in Table F
NP Chunking VGChunking
Recall
Precision
Table F Text Chunking p erformance of the XTAGparser
System Training Size Recall Precision
Ramshaw Marcus Baseline
Ramshaw Marcus
without lexical information
Ramshaw Marcus
with lexical information
Sup ertags Baseline
Sup ertags
Sup ertags
Table F Performance comparison of the transformation based noun chunker and the sup ertag
based noun chunker
As describ ed earlier the results cannot b e directly compared with other results in chunking
such as in Ramshaw and Marcus since we do not train from the Treebank b efore testing
However in earlier work text chunking was done using a technique called sup ertagging Srinivas
b which uses the XTAG English grammar which can b e used to train from the Treebank
The comparative results of text chunking b etween sup ertagging and other metho ds of chunking
is shown in Figure F
We also p erformed exp eriments to determine the accuracy of the derivation structures pro
duced by XTAG on WSJ text where the derivation tree pro duced after parsing XTAG is
interpreted as a dep endency parse Wetooksentences that were words or less from the Penn
Treebank Marcus et al The sentences were collected from sections of the Tree
AG bank of these sentences were given at least one parse by the XTAG system Since XT
typically pro duces several derivations for each sentence we simply picked a single derivation
We treat a sequence of verbs and verbal mo di ers including auxiliaries adverbs mo dals as constituting a
verb group
It is imp ortant to note in this comparison that the sup ertagger uses lexical information on a p er word basis
only to pick an initial set of sup ertags for a given word
from the list for this evaluation Better results mightbeachieved by ranking the output of the
parser using the sort of approach describ ed in Srinivas et al
There were some striking di erences in the dep endencies implicit in the Treebank and those
given byXTAG derivations For instance often a sub ject NP in the Treebank is linked with the
rst auxiliary verb in the tree either a mo dal or a copular verb whereas in the XTAG derivation
the same NP will b e linked to the main verb Also XTAG pro duces some dep endencies within
an NP while a large number of words in NPs in the Treebank are directly dep endent on the
verb To normalize for these facts we to ok the output of the NP and VG chunker describ ed
ab ove and accepted as correct any dep endencies that were completely contained within a single
chunk
For example for the sen tence Borrowed shares on the Amex rose to another record the
XTAG and Treebank chunks are shown b elow
XTAG chunks
Borrowed shares on the Amex rose
to another record
Treebank chunks
Borrowed shares on the Amex rose
to another record
Using these chunks we can normalize for the fact that in the dep endencies pro duced by
XTAG borrowed is dep endenton shares i e in the same chunk while in the Treebank borrowed
is directly dep endent on the verb rose That is to say we are lo oking at links b et ween chunks
not b etween words The dep endencies for the sentence are given b elow
XTAG dependency Treebank dependency
Borrowed shares Borrowed rose
shares rose shares rose
on shares on shares
the Amex the Amex
Amex on Amex on
rose NIL rose NIL
to rose to rose
another record another record
record to record to
After this normalization testing simply consisted of counting how many of the dep endency
links pro duced by XTAG matched the Treebank dep endency links Due to some tokenization
and subsequent alignment problems we could only test on of the original parsed
sentences There were a total of dep endency links extracted from the Treebank The
XTAG parses also pro duced dep endency links for the same sentences Of the dep endencies
pro duced bytheXTAG parser were correct giving us an accuracy of
APPENDIX F EVALUATION AND RESULTS
F Comparison with IBM
The evaluation in this section was done with the earlier release of the grammar This
section describ es an exp eriment to measure the crossing bracket accuracy of the XTAG parsed
IBM manual sentences In this exp eriment XTAG parses of IBM manual sentences have
been ranked using certain heuristics The ranked parses have been compared against the
bracketing given in the Lancaster Treebank of IBM manual sentences Table F shows the
results of XTAG obtained in this exp eriment which used the highest ranked parse for each
system It also shows the results of the latest IBM statistical grammar Jelinek et al
on the same genre of sentences Only the highest ranked parse of b oth systems was used for this
evaluation Crossing Brackets is the p ercentage of sentences with no pairs of brackets crossing
the Treebank bracketing i e a b c has a crossing bracket measure of one if compared
to a b c Recall is the ratio of the number of constituents in the XTAG parse to the
Precision is the ratio of the number of constituents in the corresp onding Treebank sentence
numb er of correct constituents to the total numb er of constituents in the XTAG parse
System of Crossing Bracket Recall Precision
sentences Accuracy
XTAG
IBM Statistical
grammar
Table F Performance of XTAG on IBM manual sentences
As can be seen from Table F the precision gure for the XTAG system is considerably
lower than that for IBM For the purp oses of comparative evaluation against other systems
we had to use the same crossing brackets metric though we b elievethat the crossing brackets
measure is inadequate for evaluating a grammar like XTAG There are two reasons for the
inadequacy First the parse generated by XTAG is much richer in its representation of the
internal structure of certain phrases than those present in manually created treebanks e g
IBM your p ersonal computer XTAG your p ersonal computer This
N NP G N N N
is re ected in the numb er of constituents p er sentence shown in the last column of Table F
System Sent of Av of Av of
Length sent words sent Constituents sent
XTAG
IBM Stat
Grammar
Table F Constituents in XTAG parse and IBM parse
A second reason for considering the crossing bracket measure inadequate for evaluating
We used the parseval program written by Phil Harison phil atc b o eing com
The Treebank was obtained through Salim Roukos roukos watson ibm com at IBM
We are aware of the fact that increasing the number of constituents also increases the recall p ercentage
However we b elieve that this a legitimate gain
XTAG is that the primary structure in XTAG is the derivation tree from which the bracketed
tree is derived Two identical bracketings for a sentence can have completely di erent derivation
trees e g kick the bucket as an idiom vs a comp ositional use A more direct measure of the
p erformance of XTAGwould evaluate the derivation structure which captures the dep endencies
between words
F Comparison with Alvey
The evaluation in this section was done with the earlier release of the grammar This
section compares XTAG to the Alvey Natural Language To ols ANLT Grammar We parsed
the set of LDOCE Noun Phrases presented in App endix B of the technical rep ort Carroll
using XTAG Table F summarizes the results of this exp erimen t Atotal of noun
phrases were parsed The NPs which did not have a correct parse in the top three derivations
were considered failures for either system The maximum and average number of derivations
columns show the highest and the average number of derivations pro duced for the NPs that
have a correct derivation in the top three We show the p erformance of XTAG b oth with and
without the tagger since the p erformance of the POS tagger is signi cantly degraded on the
NPs b ecause the NPs are usually shorter than the sentences on whichitwas trained It would
be interesting to see if the two systems p erformed similarly on a wider range of data
System of parsed parsed Maximum Average
NPs derivations derivations
ANLTParser
XTAGParser with
POS tagger
XTAGParser without
POS tagger
Table F Comparison of XTAG and ANLTParser
F Comparison with CLARE
The evaluation in this section was done with the earlier release of the grammar This
section compares the p erformance of XTAG against that of the CLARE system Alshawi
et al on the ATIS corpus Table F shows the p erformance results The p ercentage
parsed column for b oth systems represents the p ercentage of sentences that pro duced any parse
It must b e noted that the p erformance result shown for CLARE is without any tuning of the
grammar for the ATIS domain The p erformance of CLARE a later version of the CLARE
system is estimated to b e higher than that of the CLARE system
In an attempt to compare the p erformance of the two systems on a wider range of sentences
from similar genres we provide in Table F the p erformance of CLARE on LOB corpus and
When CLARE is tuned to the ATIS domain p erformance increases to However XTAG has not b een
tuned to the ATIS domain
APPENDIX F EVALUATION AND RESULTS
System Mean length parsed
CLARE
XTAG
Table F Performance of CLARE and XTAGontheATIS corpus
the p erformance of XTAG on the WSJ corpus The p erformance was measured on sentences of
up to words for b oth systems
System Corpus Mean length parsed
CLARE LOB
XTAG WSJ
Table F Performance of CLARE and XTAG on LOB and WSJ corpus resp ectively
Bibliography
Ab eill eandSchab es Anne Ab eill e and Yves Schab es Parsing Idioms in Lexicalized
TAGs In Proceedings of EACL pages
Ab eill e Anne Ab eill e French and english determiners Interaction of morphology
syntax and semantics in lexicalized tree adjoining grammars In Tree Adjoining Gram
mar First International Workshop on TAGs Formal Theory and Applications abstracts
Shloss Dagstuhl Sarrbruc ken
Abney Steven Abney The English Noun Phrase in its Sentential Aspects PhD thesis
MIT
Abney Steven Abney Parsing bychunks In Rob ert Berwick Steven Abney and Carol
Tenny editors Principle based parsing Kluwer Academic Publishers
Akma jian A Akma jian On deriving cleft sentences from pseudo cleft sentences Lin
guistic Inquiry
Alshawi et al Hiyan Alshawi David Carter Richard Crouch Steve Pullman Manny
Rayner and Arnold Smith CLARE A Contextual Reasoning and Cooperative Response
Framework for the Core Language Engine SRI International Cambridge England
Ball Catherine N Ball The historical development of the it cleft PhD thesis University
of Pennsylvania
Baltin Mark Baltin Heads and pro jections In Mark Baltin and Anthony S Kro ch
editors Alternative conceptions of phrase structure pages University of Chicago Press
Chicago Illinois
Barwise and Co op er John Barwise and Robin Co op er Generalized Quanti ers and
Natural Language Linguistics and Philosophy
Becker Tilman Becker HyTAG A new Type of Tree Adjoining Grammars for Hybrid
at des Saar Syntactic Representation of Free Word Order Languages PhD thesis Universit
landes
Becker Tilman Becker Patterns in metarules In Proceedings of the rd TAG Confer
ence Paris France
Bhatt Ra jesh Bhatt Pro control in tags Unpublished pap er UniversityofPennsylva
nia
BIBLIOGRAPHY
Browning Marguerite Browning Nul l Operator Constructions PhD thesis MIT
Burzio Luigi Burzio Italian syntax A Government Binding approach Studies in nat
ural language and linguistic theory Reidel Dordrecht
Candito Marie Helene Candito A Principle Based Hierarchical Representation of LT
AGs In Proceedings of COLING Cop enhagen Denmark
Carroll John Carroll Practical Uni cation based Parsing of Natural Language Univer
sityofCambridge Computer Lab oratory Cambridge England
Chomsky and Lasnik Noam Chomsky and Howard Lasnik The minimalist program
manuscript
Chomsky Noam Chomsky Aspects of the Theory of Syntax MIT Press Cambridge
Massac husetts
Chomsky Noam Chomsky Remarks on Nominalization In Readings in English Trans
formational Grammar pages Ginn and Company Waltham Massachusetts
Chomsky Noam Chomsky Barriers MIT Press Cambridge Massachusetts
Chomsky Noam Chomsky A Minimalist Approach to Linguistic Theory MIT Working
Papers in Linguistics Occasional Pap ers in Linguistics No
Church Kenneth Ward Church A Sto chastic Parts Program and Noun Phrase Parser for
Unrestricted Text In nd Applied Natural Language Processing Conference Austin Texas
Cinque G Cinque Types of A bar Dependencies MIT Press Cambridge Mas
sachusetts London England
Cowie and Mackin A P Cowie and R Mackin editors Oxford Dictionary of Current
Idiomatic English volume Oxford University Press London England
Delahunty Gerald P Delahunty The Analysis of English Cleft Sentences Linguistic
Analysis
Delin Judy L Delin Cleft Constructions in Discourse PhD thesis University of Edin
burgh
Doran Christine Doran Incorporating Punctuation into the Sentence Grammar A
Lexicalized Tree Adjoining Grammar Perspective PhD thesis University of Pennsylv ania
Egedi and Martin Dania Egedi and Patrick Martin AFreely Available Syntactic Lexi
con for English In Proceedings of the International Workshop on Sharable Natural Language
Resources Nara Japan August
Emonds J Emonds Root and structure preserving transformations PhD thesis Mas
sachusetts Institute of Technology
Ernst T Ernst More on Adverbs and Stressed Auxiliaries Linguistic Inquiry pages
Evans et al Roger Evans Gerald Gazdar and David Weir Enco ding Lexicalized Tree
Adjoining Grammars with a Nonmonotonic Inheritance Hierarchy In Proceedings of the rd
Annual Meeting of the Association for Computational Linguistics Cambridge MA
Gazdar et al G Gazdar E Klein G Pullum and I Sag GeneralizedPhrase Structure
Grammar Harvard University Press Cambridge Massachusetts
Grimshaw Jane Grimshaw Argument Structure MIT Press Cambridge Massachusetts
Haegeman Liliane Haegeman Introduction to Government and Binding Theory Black
well Publishers Cambridge Massachusetts Oxford U K
Hale and Keyser Ken Hale and Samuel Jay Keyser Some transitivity alternations in
English Technical Rep ort MITCCS July
Hale and Keyser Ken Hale and Samuel Jay Keyser A view from the middle Technical
Rep ort MITCCS January
Hanks Patrick Hanks editor Col lins Dictionary of the English Language Collins
London England
Heyco ck and Kro ch Caroline Heyco ckandAnthony S Kro ch Verb movement and the
status of sub jects implications for the theory of licensing GAGL
Heyco ck Caroline Heyco ck Progress Rep ort The English Copula in a LTAG Internal
XTAG pro ject rep ort UniversityofPennsylvania Summer
Ho ckey and Mateyak Beth Ann Ho ckey and Heather Mateyak Determining determiner
sequencing Asyntactic analysis for english Technical Rep ort Institute for Research
in Cognitive Science UniversityofPennsylvania
ey and B Srinivas Feature Based TAG in Place Ho ckey and Srinivas Beth Ann Ho ck
of Multi comp onent Adjunction Computational Implications In Proceedings of the Natural
Language Processing Paci c Rim Symposium NLPRS Fukuoka Japan December
Ho ckey Beth Ann Ho ckey Echo questions intonation and fo cus In Proceedings of the
Interdisciplinary ConferenceonFocus and Natural Language Processing in Celebration of the
th
anniversary of the Journal of Semantics Eschwege Germany
Hornby A S Hornby editor OxfordAdvancedLe arner s Dictionary of Current English
Oxford University Press London England third edition
Iatridou Sabine Iatridou Topics in Conditionals PhD thesis MIT
Jackendo R Jackendo Semantic Interpretation in Generative Grammar MIT Press
Cambridge Massachusetts
BIBLIOGRAPHY
Jackendo R Jackendo On Larson s Analysis of the Double Ob ject Construction
Linguistic Inquiry
Jelinek et al Fred Jelinek John La erty David M Magerman Rob ert Mercer Adwait
Ratnaparkhi and Salim Roukos Decision Tree Parsing using a Hidden Derivation Mo del
In Proceedings from the ARPA Workshop on Human Language Technology Workshop March
Joshi et al Aravind K Joshi L Levy and M Takahashi Tree Adjunct Grammars
Journal of Computer and System Sciences
Joshi Aravind K Joshi Tree Adjoining Grammars How muc h context Sensitivity is
required to provide a reasonable structural description In D Dowty I Karttunen and
A Zwicky editors Natural Language Parsing pages Cambridge University Press
Cambridge U K
Kaplan and Bresnan Ronald Kaplan and Joan Bresnan Lexical functional Grammar
AFormal System for Grammatical Representation In J Bresnan editor The Mental Rep
resentation of Grammatical Relations MIT Press Cambridge Massachusetts
Karp et al Daniel Karp Yves Schab es Martin Zaidel and Dania Egedi A Freely
th
Available Wide Coverage Morphological Analyzer for English In Proceedings of the
International Conference on Computational Linguistics COLING Nantes France Au
gust
Keenan and Stavi E L Keenan and J Stavi ASemantic Characterization of Natural
Language Determiners Linguistics and Philosophy August
Knowles John Knowles The Cleft Sentence A Base Generated Persp ective Lingua
International Review of General Linguistics August
Kro ch and Joshi Anthony S Kro ch and Aravind K Joshi The Linguistic Relevance of
Tree Adjoining Grammars Technical Rep ort MS CIS Department of Computer and
Information Science UniversityofPennsylvania
Kro ch and Joshi AnthonyS Kroch and Aravind K Joshi Analyzing Extrap osition in
aTree Adjoining Grammar In G Huck and A Ojeda editors Discontinuous Constituents
olume Academic Press Syntax and Semantics v
Lap ointe S Lap ointe A lexical analysis of the English auxiliary verb system In
T Ho ekstra et al editor Lexical Grammar pages Foris Dordrecht
Larson R Larson On the Double Ob ject construction Linguistic Inquiry
Larson R Larson Double Ob jects Revisited Reply to Jackendo Linguistic Inquiry
Lasnik and Saito H Lasnik and M Saito On the Nature of Prop er Government Lin
guistic Inquiry
Lasnik and Uriagereka H Lasnik and J Uriagereka A Course in GB Syntax MIT
Press Cambridge Massachusetts
Lees Rob ert B Lees The Grammar of English Nominalizations Indiana University
Research Center in Anthrop ology Folklore and Linguistics Indiana University Blo omington
Indiana
Lehmann et al Sabine Lehmann Stephan Oep en Sylvie Regnier Prost Klaus Netter
Veronika Lux Judith Klein Kirsten Falkedal Frederik Fouvry Dominique Estival Eva
Dauphin Herv e Compagnion Judith Baur Lorna Balkan and Doug Arnold tsnlp
Test Suites for Natural Language Pro cessing In Proceedings of COLING Kop enhagen
Lib erman Mark Lib erman Tex on tap the ACL Data Collection Initiative In Proceed
ings of the DARPA Workshop on Speech and Natural Language Processing Morgan Kaufman
Marcus et al Mitchell M Marcus Beatrice Santorini and Mary Ann Marcinkiewicz
Building a Large Annotated Corpus of English The Penn Treebank Computational Lin
guistics June
Mateyak Heather Mateyak Negation of noun phrases with not Technical Rep ort
UPENN IRCS Octob er
McCawley JamesD McCawley The Syntactic Phenomena of English The University
of Chicago Press Chicago Illinois
Moro A Moro There as a Raised Predicate Pap er presented at GLOW
Nap oli Donna Jo Nap oli Ergative Verbs in English Journal of the Linguistic Society
of America Language March
Ob ernauer H Ob ernauer On the Identi cation of Empty Categories Linguistic Review
Paroub ek et al PatrickParoub ek Yves Schab es and Aravind K Joshi Xtag a graph
ical workb ench for developing tree adjoining grammars In Third Conference on Applied
Natural Language Processing Trento Italy
Partee et al Barbara Partee Alice ter Meulen and Rob ert E Wall Mathematical
Methods in Linguistics Kluwer Academic Publishers
Pollard and Sag Carl Pollard and Ivan A Sag Head Driven Phrase StructureGrammar
CSLI
Pollo c k J Y Pollo ck Verb Movement UG and the Structure of IP Linguistic Inquiry
Quirk et al Randolph Quirk Sidney Greenbaum Geo rey Leech and Jan Svartvik A
Comprehensive Grammar of the English Language Longman London
BIBLIOGRAPHY
Ramshaw and Marcus Lance Ramshaw and Mitchell P Marcus Text chunking using
transformation based learning In Proceedings of the Third Workshop on Very Large Corpora
MIT Cambridge Boston
Rizzi Luigi Rizzi Relativized Minimality MIT Press Cambridge Massachusetts Lon
don England
Rogers and Vijay Shankar James Rogers and K Vijay Shankar Obtaining Trees from
their Descriptions An Application to Tree Adjoining Grammars Computational Intel ligence
Rosenbaum Peter S Rosenbaum The grammar of English predicate complement con
structions MIT press Cambridge Massachusetts
Sag et al I Sag G Gazdar T W asow and S Weisler Co ordination and How to
distinguish categories Natural Language and Linguistic Theory
Sarkar and Joshi Ano op Sarkar and Aravind Joshi Co ordination in Tree Adjoining
th
Grammars Formalization and Implementation In Proceedings of the International
Conference on Computational Linguistics COLING Cop enhagen Denmark August
Schab es and Joshi Yves Schab es and Aravind K Joshi An Early Typ e Parsing Algo
th
ciation rithm for Tree Adjoining Grammars In Proceedings of the Meeting of the Asso
for Computational Linguistics Bu alo June
Schab es et al Yves Schab es Anne Ab eill e and Aravind K Joshi Parsing strategies
with lexicalized grammars Application to Tree Adjoining Grammars In Proceedings of
th
the International Conference on Computational Linguistics COLING Budap est
Hungary August
Schab es Yves Schab es Mathematical and Computational Aspects of Lexicalized Gram
mars PhD thesis Computer Science Department UniversityofPennsylvania
Seuss Dr Seuss The Lorax Random House New York New York
ong Huang Fast Tree Trellis Search for So ong and Huang Frank K So ong and Eng F
Finding the N Best Sentence Hyp othesis in Continuous Sp eech Recognition Journal of
Acoustic Society AM May
Sornicola R Sornicola IT clefts and WH clefts twoawkward sentence typ es Journal
of Linguistics
Srinivas et al B Srinivas D Egedi C Doran and T Becker Lexicalization and gram
mar development In Proceedings of KONVENS pages Vienna Austria
Srinivas et al B Srinivas Christine Doran and Seth Kulick Heuristics and parse rank
th
chnologies ing In Proceedings of the Annual International Workshop on Parsing Te
Prague Septemb er
Srinivas a B Srinivas Complexity of Lexical Descriptions and its Relevance to Partial
Parsing PhD thesis UniversityofPennsylvania Philadelphia PA August
Srinivas b B Srinivas Performance Evaluation of Sup ertagging for Partial Parsing In
Proceedings of Fifth International Workshop on Parsing Technology Boston USA September
Vijay Shanker and Joshi K Vijay Shanker and Aravind K Joshi Uni cation Based
Tree Adjoining Grammars In J Wedekind editor Uni cation basedGrammars MIT Press
Cambridge Massachusetts
Vijay Shanker and Schab es K Vijay Shanker and Yves Schab es Structure sharing in
th
lexicalized tree adjoining grammar In Pr oceedings of the International Conference on
Computational Linguistics COLING Nantes France August
Watanab e Akira Watanab e The Notion of Finite Clauses in AGR Based Case Theory
MIT Working Papers in Linguistics
XTAG Group The XTAG Group A Lexicalized Tree Adjoining Grammar for English
Technical Rep ort IRCS UniversityofPennsylvania