Language use in context

Janyce Wieb e Graeme Hirst and Diane Horton

THIS IS NOT THE FINAL VERSION

Authors addresses

Wieb e Department of Computer Science and the Computing Research Lab oratory New

Mexico State University Box Dept CS Las Cruces NM email wieb ecsnmsuedu

phone fax

Hirst Department of Computer Science University of Toronto Toronto Ontario Canada

fax MS A email ghcstorontoedu phone

Department of Computer Science University of Toronto Toronto Ontario Canada Horton

MS A email dianehcstorontoedu phone fax

Despite the new emphasis on applicationdriven research in computational linguistics much

fundamental research remains to b e done In this pap er we describ e recent computational

work investigating language use in context This means rst going b eyond sentence b ound

oursetreating texts or dialogues as wholes comp osed of interrelated aries and pro cessing disc

parts rather than merely as sequences of isolated sentences Any piece of discourse establishes

a linguistic context against which subsequent utterances must b e understo o d Second b eyond

articipatory context A sp eaker or writer directs an utterance or the linguistic context is the p

text toward a hearer or reader and do es so with a particular intent or purp oseto inform to

amuse to collab orate in a task p erhaps The form and content of the utterance are chosen

accordingly and the listener or reader must infer this intent as part of understanding

We will discuss the comprehension and pro duction of language lo oking at b oth texts and

dialogues A text to b e pro cessed might b e for example a newspap er or magazine article

o d or that is b eing translated into another language or whose content is to b e understo

gue to b e pro cessed might abstracted in an information storage and retrieval system A dialo

b e a conversation sp oken or typ ed b etween a human and a computer in service of some

collab orative task Many of the problems that we shall describ e b elow o ccur in b oth kinds

of discourse We will use the terms speaker and writer almost interchangeably and similarly

he arer and reader

The underlying goal of the research describ ed in this sp ecial issue is to move b eyond

toy systems and come to grips with real language While the research describ ed in

the other pap ers in this issue fo cuses on robustly pro cessing massive amounts of text the

work describ ed here fo cuses on understanding in computational terms the complexities and

subtleties of language as p eople really use it

In a pap er of this length we cannot hop e to describ e all of the recent imp ortant work

addressing language use in context For example we will not discuss work on pronoun reso

lution ellipsis metaphor or many asp ects of b elief ascription

STR UCTURE BEYOND THE SENTENCE BOUNDARY

Discourse segmentation

Discourse has a rich structure Sentences group together and there are a variety of ways in

which they might b e related to one another Understanding what a discourse means requires

determining how the various pieces t together Consider the following excerpt from the

intro duction to a textb o ok on programming in C

{ One of the central goals of this text is to enable teachers to manage Cs inherent

complexity

Managing complexity however is precisely what we do as programmers

When we are faced with a problem that is to o complex for immediate solution

we divide it into smaller pieces and consider each one indep endently

Moreover when the complexity of one of those pieces crosses a certain threshold

it makes sense to isolate that complexity by dening a separate abstraction that

has a simple interface

The interface protects clients from the underlying details of the abstraction thereby

simplifying the conceptual structure

{ The same approach works for teaching programming

To make the material easier for students to learn this text adopts a librarybased

1

approach that emphasizes the principle of abstraction

When a p erson reads this excerpt he or she gets more from it than just the meanings of

e relationships among the individual sentences Understanding the rhetorical or coherenc

pieces of the text is an imp ortant asp ect of understanding the text as a whole For example

sentences give sp ecic details that expand on what is said in sentence Moreover

text has relationships at many levels for example not only are sentences related to

sentence but they have relationships among themselves as well Fortunately there are

sometimes cue phrases or textual markings that help the reader gure out the relationships

in sentence for and in sp oken dialogue intonation can help to o The word however

instance signals some sort of contrast with what was said in

Readers have to recover the structure of a discourse not only so that they can infer the rela

tionships among the pieces but also b ecause the structure constrains other essential asp ects of

understanding such as guring out what a pronoun refers to Consider this conversation

A Sheila wants you to call her ab out the bicycle

Has she found a ro ommate yet B

A Yeah

Her old friend Linda is moving here from Waterlo o to start a new job

She moves in next week

Anyway you should phone her to day

The pronoun her in sentence clearly refers to Sheila even though it was Linda who was

just referred to by she in the previous sentence The structure of the dialogue can help the

reader to identify the correct referent Sentences interrupt what was under discussion

in but sentence returns to the topic of Notice that the cue phrase anyway

p ossibly accompanied by a small pause or change in pitch patterns gives a strong hint of this

structure The structure constrains the p ossible referents for the pronoun her in the

referent is more likely to come from sentence than from In fact the p erception

of the structure of the discourse and the interpretation of pronouns constrain one another

In addition to relationships concerning the content of the text there are relationships

concerning the writers or sp eakers intentions and the two are closely linked In example

the author had a reason for writing sentences p erhaps to clarify for the reader

what was written in The details given in serve this purp ose of the author

There is evidence that p eople do p erform this kind of segmentation of discourse during

understanding For example P assonneau and Litman and Hirschb erg and Grosz found

1

Eric S Rob erts The Art and Science of C A LibraryBased Introduction to Computer Science Addison

Wesley page xv

statistically signicant agreement among sub jects who were asked to p erform a discourse

segmentation task How might a computer p erform such segmentation or pro duce language

from which such segments can b e recovered In their study Hirschb erg and Grosz also

investigated the relationship b etween features of intonation such as pitch range and timing

and the structure of the discourse They found that at b oth the lo cal and global levels of

discourse there are statistically signicant correlations b etween certain features of discourse

structure and certain intonational features

Another imp ortant kind of indication is cue phrases such as anyway however wel l stil l

example and now which often provide explicit information ab out the structure of a for

discourse For example now can b e used to intro duce a new subtopic But many words

that serve as cue phrases also have other usesfor example now can simply mean at this

time To see how these ambiguities could b e resolved Hirschb erg and Litman studied

cuephrase usage in sp eech and develop ed a metho d for disambiguating cue phrases by means

of intonational features of sp eech They also discovered textual features of transcrib ed sp eech

such as punctuation that are relatively easy to extract from transcriptions and can b e used

as additional aids in cuephrase disambiguation Much computational work in discourse

pro cessing has assumed that cuephrase disambiguation is feasible Hirschb erg and Litmans

ndings provide empirical supp ort for this assumption

The research describ ed ab ove has obvious applications in b oth sp eech generation and

sp eech understanding Sp eech generation systems can use intonation and cue phrases as p eo

ple do to help to break the discourse into appropriate segments thereby assisting the listener

in accomplishing other tasks that are crucial to understanding sp eech such as determining the

referents of noun phrases and recognizing the rhetorical and intentional relationships b etween

segments And sp eech recognition systems can use intonational features and cue phrases and

p erhaps also textual features if the sp eech has b een transcrib ed to help p erform segmentation

and infer relationships b etween segments In fact ATT Bell Lab oratories TexttoSp eech

System based on Hirschb erg and Litmans work do es b oth it disambiguates cue phrases

in text on the basis of textual features and then generates cue phrases in such a way that

the intonational features suggest how the cue phrases are b eing used

Relationships within discourse segments

We now turn to some recent work that investigates the structure within discourse segments

Hobbs et al and others have suggested that during understanding p eople make de

feasible assumptions that is assumptions that are consistent with what they b elieve but

which can b e later overridden by contrary evidence Such assumptions lead them to a plau

y and Jensen apply this approach sible coherent interpretation of the discourse Zadrozn

holistically to paragraphs with the goal of nding interpretations of individual sentences in

a paragraph that are all together consistent Their approach formalizes the intuitive notion

that the sentences in a paragraph all are related in some way to the same topic

Lascarides Asher and Ob erlander whose approach is also founded on defeasible

reasoning address the inference of certain typ es of coherence relations b etween segments

and of temporal relations b etween the events that the discourse refers to In the absence of

particular information to the contrary the default coherence relation b etween two sentences

and s and e is simple narration in which case e that describ e resp ectively events e s

1 2 1 2 1

o ccurs b efore e In for example one assumes that Johns closing the do or precedes his

2

sitting down on the couch

John closed the do or to the kitchen

He sat down on the couch

But in

John fell

Max had pushed him

we infer the stronger relation that Maxs pushing John caused John to fall in which case

the second event precedes the rst one notice that the tenses in supp ort this temp oral

interpretation the coherence relation in this case is that is an explanation of

But p eople are sometimes sloppy in their use of tense more technically tense and aspect

one might have come to the same conclusion even if b oth sentences were in the simple past

b ecause of ones knowledge ab out pushing and falling

John fell

Max pushed him

Fo cusing only on background knowledge and ignoring tense Lascarides et al express defaults

such as that describ ed ab ove for narration in a nonmonotonic logic These are overridden

if there is information to the contrary in for example the default is overridden by

the sp ecic knowledge that pushing someone causes them to fall Their mechanism is also

sensitive to the linguistic context one could imagine a context for in which it is taken to

mean that John fell and then Max pushed him In this case the relation is narration after

all

Hwang and Schub ert fo cus on interpreting tense to create a representation of the

temp oral relations among events describ ed in the discourse including implicit temp oral re

lations across clause and sentence b oundaries To get an idea of what sorts of relations are

recognized consider from

John went to the hospital

The do ctor told John that he had broken his ankle

Hwang and Schub erts mechanism derives the following relations among others Johns going

to the hospital to ok place b efore the moment at which is said the do ctor informing

John of something happ ened after Johns going to the hospital but b efore the time when the

sentences in are said and Johns breaking his ankle to ok place b efore the do ctors telling

him he broke it

Hwang and Schub erts mechanism also works for longer narrative provided that tense is

used literally Consider the following mo died version of an example given in The

refers to the time of event e notation t

e

John went over to Marys house t

oO v er g

wer shop for some roses t On the way he had stopp ed by the o

stop

He had picked out ve red ones three white ones and one pale pink t

pick O ut

Then he had chosen a vase to put them in t

choose

Unfortunately they failed to cheer her up t

f ailT oC heer

Tense technically tense asp ect and the aspectual classes of the events and states implies

the following temp oral relations time t is b efore the time at which is said since

g oO v er

is in the simple past The past p erfect tense of had stopped in had picked out

in and had chosen in implies that t t and t are b efore the end

stop pick O ut choose

of t There are other relations p ossible with the past p erfect but this issue will not

g oO v er

concern us here Further with the simple past tense of failed to cheer her up in we

return to the time when John is at Marys house Sentences thus form a subnarrative

emb edded within the overall narrative

Of course tenses are not always used as literally as in It is quite natural for the simple

past rather than the past p erfect to b e used once p ersp ective is shifted to the subnarrative

John went over to Marys house t

g oO v er

On the way he had stopp ed by the ower shop for some roses t

stop

He picked out ve red ones three white ones and one pale pink t

pick O ut

Then he chose a vase to put them in t

choose

Unfortunately they failed to cheer her up t

f ailT oC heer

Some of the tenses are dieren t in and yet the temp oral relations among events are the

same Notice that are in the simple past like yet describ e events in the

emb edded narrative but resumes the main narrative Thus as Hwang and Schub ert and

others discuss tense alone is not sucien t in such cases to determine the temp oral relations

among events

Kamey ama Passonneau and Po esio take up this problem Their approach is based

on the idea prop osed by Webb er among others that determining the time that a past

tense refers to is similar to determining which entity a pronoun refers to in that they b oth

dep end on things mentioned in the previous discourse In b elow the time that the past

dep ends on the event describ ed in the previous sentence g etO n in tense refers to t

f f r ideO

particular t is after t whenever t might b e

r ideO f f g etO n g etO n

The ranger got on his horse t

g etO n

He ro de o into the sunset t

r ideO f f

In Kameyama et als terminology a past tense is understo o d with resp ect to a discourse

reference time which is established by the linguistic context For the second sentence the

discourse reference time is t Now consider from where a and b

g etO n

are two alternative continuations

John went over to Marys house t

g oO v er

On the way he had stopp ed by the ower shop for some roses t

stop

a He picked out ve red ones three white ones and one pale pink t

pick O ut

Unfortunately they failed to cheer her up t b

f ailT oC heer

Both continuations are in the simple past Yet the rst is part of the emb edded narrative

while the second is part of the main narrative In Kameyama et als analysis the past p erfect

had stopped intro duces two discourse reference times and a following pasttense sentence

might b e understo o d with resp ect to either one of them The two times intro duced by

are t actually a time inferred to b e equal to the end of t and t the past tense

g oO v er g oO v er stop

of a is understo o d with resp ect to t while the past tense of b is understo o d

stop

with resp ect to the end of t Kameyama et al discuss how to keep track of discourse

g oO v er

reference times as the discourse pro ceeds and cho osing the right one for a given past tense

Another typ e of segmentation

In texts writers often rep ort the mental states b eliefs knowledge intentions hatred p ercep

tions and so on of many dierent p eople The most straightforward way to rep ort someones

mental state is to present it explicitly as such with a sentence such as

Stuart had accomplished his mission

But he knew that by now the enemy was swarming in his rear

To return the way he had come would invite trouble

To continue on to make a complete circuit around McClellans army might

foil the pursuit

2

Besides it would b e a glorious achievement

But mental states may also b e presented implicitly as in Example is a passage

from a nonction b o ok ab out the American Civil War in fact Stuart did not turn back

but continued on In the writer is presenting Stuarts motivations for doing so

even though the writer do es not explicitly indicate that he is presenting them Thus we

have the problem of segmenting text according to whose b eliefs intentions and so forth are

b eing presented a problem made dicult by such implicitly presented mental states Wieb e

develop ed an algorithm for p erforming this segmentation in thirdp erson narrative texts

The algorithm is based on regularities found by extensive examination of naturally o ccurring

text in the ways that writers manipulate p oint of view For example an explicit rep ort

of an agents mental state can b e an indication that a blo ck of sentences presenting that

agents mental states will follow as in But this indication do es not typically hold if

the sentence contains an expression of uncertainty or judgment toward the mental state such

textual markings suggest the p oint of view of either another p erson mentioned in the text or

the writer Thus expressions of uncertainty or judgement are similar to cue phrases discussed

earlier in this pap er b ecause they can help the reader p erform p ointofview segmentation

An example is the phrase It was almost as if in

2

James M McPherson Battle Cry of Freedom Oxford University Press page

It was almost as if he Brown knew that failure with its ensuing martyrdom would

do more to acheive his ultimate goal than any success could have done

3

In any event that was how matters turned out

Even though and b oth mention a p ersons knowledge the hedge in

suggests that Browns p oint of view do es not continue in as Stuarts do es in

Sentence do es not present Browns mental state but rather describ es the eventual

historical outcome The absence or presence of a phrase such as the hedge in is one

textual feature that can help a text understanding system recognize implicitly presented

mental states

THE SPEAKER AND THE HEARER

We now move to recent research that acknowledges the role played in discourse by the indi

vidual knowledge goals exp eriences etc of the sp eaker and hearer or writer and reader

Explanation

In advisory dialoguesdialogues in which an exp ert advises someone on how to assemble some

++

device for example or improve his or her C programsadvisees may lack the knowledge or

exp erience necessary to fully understand the exp erts explanations In such cases they often

ask followup questions A computer system playing the role of the exp ert should b e able

to participate in a dialogue with the user providing justications for its recommendations

descriptions of its problemsolving strategies and denitions of the terms it used Mo ore and

Paris have develop ed a text planner for advisory dialogues that has these capabilities

Mo ore and Paris integrate two main approaches to discourse in their work The rst seen

earlier in this pap er fo cuses on the rhetorical or coherence relationships among the segments

of the discourse the other intentional approach fo cuses on the intentions that motivate

sp eakers utterances and on the relationships among them With the intentional approach

generation is cast as the pro cess of planning a sequence of utterances that will achieve ones

goals and understanding is cast as the pro cess of inferring the sp eakers intentions from his or

her utterances Most theories of discourse include b oth intentional and rhetorical knowledge

to some extent Mo ore and Paris fo cus on explicitly representing and using b oth kinds during

pro cessing

The rhetorical knowledge in the system which is based on Mann and Thompsons Rhetor

ical Structure Theory consists of strategies for achieving communicative goals by means

of establishing rhetorical links b etween discourse segments For example one way to achieve

the goal of enabling the user to identify an ob ject is to contrast it with an ob ject that is

already known to the user using the contrast relation another is to tell the user some of the

attributes of the ob ject using the elaborationattribute relation Thus there might b e more

than one strategy that can achieve a particular goal as well a single strategy may b e used

in service of more than one typ e of goal

3

ibid page

Mo ore and Pariss system maintains a record of why it said what it said thus when the

user indicates that an explanation was not completely understo o d it can determine which of

its goals failed and attempt to achieve it again using a dierent strategy Consider from

System What characteristics of the program would you like to enhance

Readability and maintainability User

X with SETF X System You should replace SETQ

SETQ can only b e used to assign a value to a simplevariable

In contrast SETF can b e used to assign a value to any generalized

variable

A generalizedvariabl e is a storage lo cation that can b e named by any

accessor function

User What is a generalized variable

System For example the car of a cons is a generalized variable named by the

access function CAR and the cdr of a cons is a generalized variable

named by the access function CDR

The system has a goal to p ersuade the hearer to replace SETQ with SETF in order to enhance

the readability and maintainability of his or her Lisp program In utterances it

therefore motivates the replacement by describing relevant dierences b etween the ob ject

b eing replaced and the ob ject replacing it Having reasoned that the listener might not know

what a generalizedvariable is in the system explains the concept by stating its class

memb ership a storage lo cation and describing an attribute it can b e named by any accessor

function Yet in the user expresses the goal of knowing what a generalizedvariable is

The system already tried to satisfy this goal in b ecause the system explicitly represents

its own goals it can now realize that it did not succeed It therefore selects an alternative

strategy and tries again in it gives examples of generalizedvariables

Collab oration in discourse

Although the work on explanation just describ ed takes the hearer into account in some ways

discourse is fundamentally collab orative in more ways than these The participants in a

discourse work together to satisfy various of their individual and joint needs page

Most early work on inferring the intentional structure b ehind discourse did not consider this

Many theories mo deled only situations in which one agent p erforms actions b oth linguistic

utterancesand actions in the domain of discourse while the system uses these actionsie

actions as a basis for attributing intentions or plans to the agent and is otherwise passive

In many cases the system had no plans of its own nor did it consider that the agent might

attempt to attribute any plans to it There certainly were no joint plans Furthermore the

agents plans were presumed to b e preformulated All of these assumptions are incompatible

with the collab orative nature of discourse In truth all agents involved in a discourse may

have plans they may all infer each others plans they often share joint plans eg that one

will explain something to another and their plans may b e formulated on the yindeed

the need to formulate a plan may b e precisely why a user consults a system Some recent

work on plan inference has attempted to embrace these imp ortant facets of real discourse

If one wishes to mo del situations in which agents share joint plans one must ask what it

means for two agents to have a joint plan to do something Grosz and Sidner are concerned

with precisely this question They extend Pollacks earlier denition of having a plan

which was designed only for singleagent plans to the case of joint plans We paraphrase

their denition as follows

Definition Two agents have a shared plan to do action A if and only if for each subaction

involved in doing A

they mutually b elieve

a that the subaction relates in a particular way to A

b that one of them can do that action

c that he intends to do it and

d that he intends by doing it to accomplish A

and the agent who is to do the action

a do es in fact intend to do it and

b intends by doing it to accomplish A

While a shared plan is under construction the two agents will hold only some of these

intentions and mutual b eliefs At such times they are said to have a partial shared plan Of

course there are a myriad of unresolved philosophical issues that such a denition raises but

Grosz and Sidners denition is a go o d starting p oint for discussion

The dicult question still remains how do the agents acquire the intentions and very

strong mutual b eliefs required to have a shared plan This is the problem that Lo chbaum

Grosz and Sidner tackle First however they mo dify the denition in two imp ortant

ways Clause a is generalized to say that the agents have a recip e for doing A A recip e

for A although dened formally can b e thought of simply as a way of doing A This change

means that one denition suces for any sort of shared plan and also since their notion of

a recip e allows for any level of detail that agents may have a shared plan with any level of

detail Clauses d and b are mo died to require not that the subaction accomplishes

A but merely that it contributes somehow to A the notion of contribution is dened as the

transitive closure of a numb er of basic relations b etween actions By not requiring that the

subaction contribute in any one particular way the denition p ermits agents to have a shared

plan that is vague in this regard These mo dications p ermit agents to share joint plans that

are vague or lacking in detail which is certainly imp ortant during plan construction when

the plan has not yet b een fully rened

Lo chbaum Grosz and Sidner oer an algorithm for inferring the mutual b eliefs expressed

in their improved versions of clauses a and d The key part of the algorithm says that

in a context where two agents have a shared plan to do A if one of them says something

ab out some action of typ e then the hearer may conclude that the sp eaker b elieves that

contributes somehow to A further if the hearer can recognize some way in which actually

do es contribute to A he may conclude that they mutually b elieve it contributes

The trains pro ject led by Allen and Schub ert also embraces the collab orative nature

of discourse The pro jects aim is to build a system that acts as a planning assistant collab

orating with the user to help formulate plans that will meet their goals The system must

b e able to discuss goals and to form plans incrementally as the system and user interact

Building such a system makes it necessary to create and integrate comp onents that handle

problems as varied as parsing reasoning ab out the domain and reasoning ab out the b eliefs

goals and plans currently held by the system and the user

The pro jects approach to discourse centers on the concepts of planning and plan exe

cution It also takes an intentional approach with utterances viewed as linguistic actions

hence they can b e treated in the same framework as domain actions that is planned for

executed ie sp oken or written reasoned ab out ie understo o d and so on

The trains pro ject has provided a testb ed for research on many problems in discourse

as well as in other areas such as temp oral reasoning The results of this research have b een

integrated into demonstration systems that are able to participate in interesting dialogues

such as the one from which the following excerpt was taken The domain of discourse is the

shipment of commo dities by rail

User We have to make OJ There are oranges at I and an OJ factory at B

Engine E is scheduled to arrive at I at PM Shall we ship the oranges

System Yes Shall I start loading oranges in the empty car at I

The user do es not explicitly prop ose that the oranges should b e shipp ed from I to B using

engine E Yet the system infers this and can therefore determine which oranges the user

is referring to when she says shal l we ship the oranges The system simultaneously answers

es It then uses its knowledge of the question and implicitly accepts the plan by replying Y

the domain to identify two p ossible ways to complete their nowmutual plan one is to use a

b oxcar already lo cated at I and the other is to wait and use the b oxcar that comes with the

engine that will arrive at pm The system then p oses a question to nd out whether the

rst alternative is acceptable to the user

F allibility of conversants

Most research in language understanding has assumed a somewhat idealized notion of human

linguistic abilities that is p eople are seen as faultless language pro cessors whose skills AI

research strives mightily to emulate Even the work discussed ab ove on inferring joint plans

though it brings the sp eaker and hearer into full consideration do es not address their falli

bility In fact p eople are frequently unclear and imprecise in what they say and write and

as comprehenders they frequently reach no understanding or worse a mistaken understand

ing However p eople make up for this by their exibility They are for example adept at

detecting when a misunderstanding has set a conversation awry and at saying the right thing

to correct it We now turn to research that takes a similar p ersp ective on humancomputer

conversation

Misunderstandings in conversation might o ccur for example b ecause the hearer takes an

unintended sense of an ambiguous expression b ecause the hearer do es not have the necessary

background knowledge to interpret the utterance or draw the right exp ected inferences from

it or simply b ecause of an error in typing or sp eech recognition If the result is either

no interpretation at all or several p ossible interpretations from which a choice cannot b e

made the hearer can then ask for clarication can remain silent hoping that subsequent

utterances will resolve matters or can invoke additional pro cessing to try to recover Eller

and Carb erry take the third approach the interpretation of an utterance that cannot

b e coherently integrated into the current context is relaxed by a set of heuristics Relaxation

p ermits the system to consider a somewhat unlikely shift of fo cus for instance or an imprecise

use of tense

But if a conversant nds a single reasonable alb eit erroneous interpretation the mis

understanding will manifest itself only later if at all when the conversants nd themselves

talking at cross purp oses There are thus two parts to the problem rst noticing that there

has b een an earlier misunderstandingeither by oneself or by the other conversantand

secondly generating an utterance that will repair the misunderstanding For the rst part

Eller and Carb erry suggest that if the heuristics describ ed ab ove do not serve to interpret

a problematic utterance the cause might b e an earlier misunderstanding and so apply the

heuristics to earlier utterances creating alternative contexts in which the current utterance

can then b e considered they do not address the second part McRoy and Hirst have

shown that b oth parts of the problem can b e accounted for in a mo del of conversation in

which the interpretation of an utterance is characterized as ab ductive reasoning eg given

Q and P ) Q guess that P is also true and the generation of an utterance as default

reasoning see and ab ove In this mo del conversants ab ductively form defeasible exp ec

tations as to what kind of utterance is likely to o ccur next in the conversation and use these

exp ectations to monitor for dierences in understanding When sp eakers make an utterance

they use their b eliefs ab out the discourse context ab out the other participants b eliefs and

ab out conventions of discourse to select an utterance appropriate to their goals The other

participant attempts to retrace this selection pro cess ab ductively trying to identify the goal

exp ectation or misunderstanding that might have lead the sp eaker to pro duce it If McRoy

and Hirsts mo del nds more than one p ossibility it makes a choice at random the mo del

not accounting for diering likeliho o ds of the various interpretations If a misunderstand

ing on either side self or other is found a conversation participant will reinterpret earlier

utterances to nd another interpretation and utter an appropriate correction

The mo del is implemented in an extension of Prolog that p erforms ab duction and default

reasoning The mo del includes b oth interpretation and generation so two copies of the

program with dierent b eliefs and knowledge can converse with one another

NUANCE AND STYLE IN LANGUAGE

The exact choice of words phrases and sentence structure all aect the precise meaning

and eect of an utterance A writer or sp eaker cho oses consciously or not such goals as

whether to b e formal or friendly p ersuasive or dismissive clear or obscure These asp ects of

an utterance are as much a part of its message as its literal meaning and any sophisticated

natural language system needs to b e sensitive to them A machine translation system for

example would b e inadequate if in translating a business letter from English to French it

preserved the literal meaning but at the same time turned a friendly letter into a threatening

one or vice versa

Indeed the selection of what is to b e said at all dep ends up on complex interp ersonal

concerns One might cho ose material that supp orts ones own p osition or that might app eal

to the listener while omitting material that undermines ones p osition or that might annoy

a listener whom one do esnt wish to oend Hovys naturallanguage generation system

p auline was the rst to account in an integrated manner for interp ersonal concerns in

linguistic nuance and in the selection of material Although pauline was impressive it had

no theoretical basis it employed a wide variety of rules that were little more than an ad ho c

collection of heuristics Subsequent research has attempted to nd moregeneral principles

for the selection of content words and syntactic structures

One particular problem is the construction of referring expressions A referring expression

is a word or phrase that a sp eaker uses to denote some particular ob ject or entity The problem

is that there are usually many ways to do this and in any given situation some are b etter

than others For example any of the following might serve to ll the space in the sentence

shown

Ill meet you near in an hour

the tree

the tall tree

the larch

the tree with the soft lightgreen needles

the tall conifer next to the old well

the larch that is ab out feet tall

the big one

it

Which of these is b est dep ends on the previous utterancesthe last two options require

a previous referring expression to act as an antecedentas well as the listeners assumed

knowledge and the circumstances of the utterance Alternative is no go o d if the listener

do esnt know enough ab out trees to identify a larch though might serve If there is

only one tree that the sp eaker could p ossibly b e referring to detailed descriptions such as

are misleading as they spuriously imply that some kind of contrast is b eing made

Dale and Reiter have develop ed metho ds for constructing referring expressions

The rst consideration is whether a pronominal reference is p ossible if not a denite noun

phrase must b e constructed Such a noun phrase must b e b oth ecient and adequate iden

tifying the referent with the least amount of information necessary to do so unambiguously

Dale uses the notion of minimal distinguishing descriptionthe smallest set of attributevalue

pairs that will serve to discriminate the referent from other entities These metho ds also take

account of the preference in language for descriptions that use socalled basic categories for

l you please take the for a walk the word animal might example in the context Wil

b e sucient to uniquely identify the thing to b e taken for a walk but dog is still the more

natural expression Reiter p oints out that using an inappropriate category or a description

more complex than necessary creates a false implicature and he presents an algorithm for

generating referring expressions that is able to avoid such situations

The problem is somewhat dierent in interactive discourse b ecause if a referring expres

sion fails to pick out a unique entity the participants can immediately try to correct it that

is they can collab orate on the task of reaching a common understanding of the reference

Heeman and Hirst have mo deled this computationally as the construction and recognition

by two agents with p ossibly diering b eliefs of plans to refer to something In this mo del the

generation of a referring expression is viewed as the construction of a plan to bring the refer

ent to the attention of the other conversant and comprehension of a referring expression is

viewed as recognition of this p ossibly faulty plan The mo del accounts for b oth pro duction

and comprehension of referring expressions thus as with the mo del of McRoy and Hirst two

copies of the program with dierent b eliefs can talk to one another negotiating a common

understanding of a referring expression

Much of the work on language generation including that on referring expressions has

sought mostly to determine the content to b e expressed For example in Dales system a

syntactic structure is chosen at random to express the content of the description eg either

the pitted olives or the olives that have been pitted But as Hovys work showed form is just

as imp ortant DiMarco and Hirst sought to correlate a writers use of various syntactic

constructions with his or her higherlevel stylistic goals with a view to ensuring that an

automatic translation retains these goals even if that requires a dierent syntactic structure

in the target language For example

Ils se livrent alors sous des dehors irr esistibles de drolerie a une lutte sournoise et

passionnee

And from b ehind a cover of irresistibly funny wit they op en re in an artful and

4

passionate battle

The underlined phrase that interrupts the main clause in was moved by the translator

to a p osition b efore the main clause in Though is quite natural in French the

movement of the phrase was necessary to prevent what would otherwise b e a somewhat

unnatural sentence in the corresp onding English To capture this kind of linguistic intuition

DiMarco and Hirst develop ed the idea of a grammar of style which correlates the syntactic

structures of a language with a set of languageindep endent stylistic goals In translation

these goals can then b e determined in the source text and used in the generation of the new

text

4

This example is from the bilingual Programme du Festival du Cinema americain Festival de Deauville

quoted by Jacqueline GuilleminFl escher Syntaxe comparee du francais et de langlais Problemes de

Editions Ophrys p who lists and explains some of the dierences in the naturalness traduction

of various structures in English and French

CONCLUSION

The common core of the research that we have discussed in this pap er is a concern for the full

complexity of language as p eople use it in particular the complexity that arises in creating

and comprehending units larger than a single sentence in deciding up on the b est word or

expression that ts the present context and intent and in accommo dating the fallibili ty of

language users Although the results of some of this work have already b een incorp orated

into application systems much of it is exploratory basic research a necessary precursor to

the development of practical systems

wledgements Ackno

We wish to thank Yorick Wilks Lisa Smith and an anonymous referee for helpful comments

on earlier drafts of this pap er The preparation of this pap er was supp orted in part by a grant

to the second author by the Natural Sciences and Engineering Research Council of Canada

References

Allen James F Schub ert Lenhart K Ferguson George Heeman Peter Hwang

Chung Hee Kato Tsuneaki Light Marc Martin Nathaniel G Miller Bradford W

Po esio Massimo and Traum David The trains Pro ject A case study in building a

of Experimental and Theoretical conversational planning agent To app ear in Journal

rticial Intel ligence A

Dale Rob ert Gener ating Referring Expressions The MIT Press Cambridge Mass

DiMarco Chrysanne and Hirst Graeme A computational theory of goaldirected style

Linguistics Septemb er in syntax Computational

Eller Rhonda and Carb erry Sandra A metarule approach to exible plan recognition

in dialogue User Modeling and UserAdapted Interaction

Grosz Barbara J and Sidner Candace L Plans for discourse In Philip R Cohen Jerry

Morgan and Martha E Pollack Eds Intentions in Communication The MIT Press

Cambridge Mass pp

Heeman Peter and Hirst Graeme Collab orating on referring expressions Computa

tional Linguistics Septemb er

Hirschb erg Julia and Grosz Barbara Intonational features of lo cal and global discourse

structure In Proceedings of the Speech and Natural Language Workshop Feb

Harriman NY DARPA Software and Intelligent Systems Technology Oce

pp

Hirschb erg Julia and Litman Diane Empirical studies on the disambiguation of cue

phrases Computational Linguistics Septemb er

Hobbs Jerry R Coherence and coreference Cognitive Science JanuaryMarch

Hobbs Jerry R Stickel Mark Martin Paul and Edwards Douglas Interpretation

meeting of the Association for Computational as ab duction In Proceedings of the th

Linguistics June Bualo NY Asso ciation for Computational Linguistics NJ

pp

Hovy Eduard H Generating Natural Language under Pragmatic Constraints Lawrence

Erlbaum Asso ciates Hillsdale NJ

Hwang Chung Hee and Schub ert Lenhart K Tense trees as the ne structure of

discourse In Proceedings of the th meeting of the Association for Computational

Linguistics June July Newark Del Asso ciation for Computational Linguistics

NJ pp

Kameyama Megumi Passonneau Reb ecca and Po esio Massimo Temp oral centering

In Proceedings of the st meeting of the Association for Computational Linguistics

June Columbus Ohio Asso ciation for Computational Linguistics NJ

pp

Lascarides Alex Asher Nicholas and Ob erlander Jon Inferring discourse relations

in context In Proceedings of the th meeting of the Association for Computational

Linguistics June July Newark Del Asso ciation for Computational Linguistics

NJ pp

Lo chbaum Karen E Grosz Barbara J and Sidner Candace L Mo dels of plans to

supp ort communication an initial rep ort In Proceedings of the th National Conference

on Articial Intel ligence July Aug Boston Mass American Asso ciation for

Articial Intelligence Calif vol pp

Mann William and Thompson Sandra Rhetorical structure theory toward a func

tional theory of text organization Text

McRoy Susan and Hirst Graeme The repair of sp eech act misunderstandings by ab

ductive inference Computational Linguistics Septemb er Probably

Mo ore Johanna and Paris Cecile Planning text for advisory dialogues capturing in

tentional and rhetorical information Computational Linguistics Decemb er

Passonneau Reb ecca and Litman Diane Intentionbased segmentation human reli

ability and correlation with linguistic cues In Proceedings of the st meeting of the

Association for Computational Linguistics June Columbus Ohio Asso ciation

for Computational Linguistics NJ pp

Pollack Martha E A mo del of plan inference that distinguishes b etween the b eliefs of

meeting of the Association for Com actors and observers In Proceedings of the th

putational Linguistics June New York NY Asso ciation for Computational

Linguistics NJ pp

Reiter Ehud Generating descriptions that exploit a users domain knowledge In Rob ert

Dale Chris Mellish and Michael Zo ck Eds Current Research in Natural Language

Generation Academic Press London pp

Sproat Richard Hirschb erg Julia and Yarowsky David A corpusbased synthesizer

In Proceedings of the International Conference on Spoken Language Processing Octo

b er Ban

Webb er Bonnie Lynn Tense as discourse anaphor Computational Linguistics

June

Wieb e Janyce M Tracking p oint of view in narrative Computational Linguistics

June

Zadrozny Wlo dek and Jensen Karen Semantics of paragraphs Computational Lin

June guistics

CR Categories and Sub ject Descriptors I Articial Intelligence Appli

cations and Exp ert SystemsNatural Language Interfaces I Articial Intelligence

Natural Language Pro cessing

Words and Phrases Discourse Collab oration Style Additional Key

Ab out the authors

JANYCE M WIEBE works in the areas of pragmatics and discourse She is an assis

tant professor in the Department of Computer Science and a principle investigator in the

Computing Research Lab oratory at New Mexico State University She received her PhD in

computer science from the State University of New York at Bualo

Authors present address Department of Computer Science New Mexico State Univer

sity Box Dept CS Las Cruces NM email wieb ecsnmsuedu

GRAEME HIRST a professor in the articial intelligence group at the University of

Toronto is the author of two b o oks in computational linguistics He is b o okreview editor of

the journal Computational Linguistics

Authors present address Department of Computer Science University of Toronto

Toronto Ontario Canada MS A email ghcstorontoedu

DIANE HORTONs research interests are in the inference of plans b eliefs and other

mental attitudes from natural language discourse She is a lecturer at the University of

Toronto

Authors present address Department of Computer Science University of Toronto

Toronto Ontario Canada MS A email dianehcstorontoedu