Towards an implementable dependency grammar
Timo JKrvinen and Pasi Tapanainen Research Unit for Multilingual Language Technology P.O. Box 4, FIN-00014 University of Helsinki, Finland
Abstract such as a context-free phrase structure gram- mar. This leads us to a situation where the Syntactic models should be descriptively ade- parsing problem is extremely hard in the gen- I quate and parsable. A syntactic description is eral, theoretical case, but fortunately parsable autonomous in the sense that it has certain ex- in practise. Our result shows that, while in gen- plicit formal properties. Such a description re- eral we have an NP-hard parsing problem, there i lates to the semantic interpretation of the sen- is a specific solution for the given grammar that tences, and to the surface text. As the formal- can be run quickly. Currently, the speed of the ism is implemented in a broad-coverage syntac- parsing system I is several hundred words per tic parser, we concentrate on issues that must I second. be resolved by any practical system that uses In short, the grammar should be empirically such models. The correspondence between the motivated. We have all the reason to believe structure and linear order is discussed. ! that if a linguistic analysis rests on a solid de- scriptive basis, analysis tools based on the the- 1 Introduction ory would be more useful for practical purposes. i The aim of this paper is to define a dependency We are studying the possibilities of using com- grammar framework which is both linguistically putational implementation as a developing and motivated and computationally parsable. testing environment for a grammatical formal- A linguistically adequate grammar is the pri- ism. We refer to the computational implemen- I mary target because if we fail to define a de- tation of a grammar as a parsing grammar. scriptive grammar, its application is less use- 1.1 Adequacy ful for any linguistically motivated purposes. In I fact, our understanding of the potential bene- A primary requirement for a parsing grammar is fits of the linguistic means can increase only if that it is descriptively adequate. Extreme dis- our practical solutions stand on an adequate de- tortion results if the mathematical properties I scriptive basis. of the chosen model restrict the data. How- Traditionally, grammatical models have been ever, this concern is not often voiced in the dis- constructed by linguists without any consider- cussion. For example, McCawley (1982, p. 92) ation for computational application, and later, notes that such a basic assumption concerning linguistic structures that "strings are more basic I by computationally oriented scientists who have than trees and that trees are available only as a first taken a parsable mathematical model and side product of derivations that operate in terms then forced the linguistic description into the of strings" was attributable to the historical ac- I model which has usually been too weak to de- cident that early transformational grammarians scribe what a linguist would desire. knew some automata theory but no graph the- Our approach is somewhere between these Ory." I two extremes. While we define the grammar One reason for computationally oriented syn- strictly in linguistic terms, we simultaneously tacticians to favour restricted formalisms is that test it in the parsing framework. What is excep- they are easier to implement. Those who began i tional here is that the parsing framework is not restricted by an arbitrary mathematical model l Demo: http://www.conezor.fi/analysers.html | !
i' to use dependency models in the 1960's largely natural language. His main work, (1959) ad- ignored descriptive adequacy in order to develop dresses a large amount of material from typo- models which were mathematically simple and, logically different languages. It is indicative of as a consequence, for which effective parsing al- Tesni~re's empirical orientation that there are gorithms could be presented. These inadequa- examples from some 60 languages, though his cies had to be remedied from the beginning, method was not empirical in the sense that he which resulted in ad hoc theories or engineering would have used external data inductively. As I solutions 2 without any motivation in the theory. Heringer (1996) points out, Tesni~re used data There have been some serious efforts to re- merely as an expository device. However, in solve these problems. Hudson (1989), for exam- order to achieve formal rigour he developed a i ple, has attempted to construct a parser that model of syntactic description, which obviously would reflecs the claims of the theory (Word stems from the non-formal tradition developed Grammar) as closely as possible. However, it since antiquity but without compromising the seems that even linguistically ambitious depen- descriptive needs. We give a brief historical I dency theories, such as Hudson's Word Gram- overview of the formal properties inherent in mar, contain some assumptions which are at- Tesni~re's theory in Section 5 before we proceed tributable to certain mathematical properties of to the implementational issues in Section 6. I an established formalism rather than imposed 1.3 The surface syntactic approach by the linguistic data 3. These kinds of unwar- ranted assumptions tend to focus the discus- We aim at a theoretical framework where we I sion on phenomena which are rather marginal, have a dependency theory that is both descrip- if a complete description of a language is con- tively adequate and formally explicit. The lat- cerned. No wonder that comprehensive descrip- ter is required by the broad-coverage parsing ! tions, such as Quirk et al. (1985), have usually grammar for English that we have implemented. been non-formal. We maintain the parallelism between the syn- tactic structure and the semantic structure in 1.2 The European structuralist our design of the syntactic description: when a i tradition choice between alternative syntactic construc- We argue for a syntactic description that is tions in a specific context should be made, the based on dependency rather than constituency, semantically motivated alternative is selected 4. I and we fully agree with Haji~ovA (1993, p. 1) Although semantics determines what kind of that "making use of the presystemic insights of structure a certain sentence should have., from classical European linguistics, it is then possi- the practical point, of view, we have a completely I ble that constituents may be dispensed with as different problem: how to resolve the syntactic basic elements of (the characterization of) the structure in a given context. Sometimes, the sentence structure." However, we disagree with latter problem leads us back to redefine the syn- the notion of "presystemic" if it is used to imply tactic structure so that it can be detected in I that earlier work is obsolete. From a descriptive the sentence s . Note, however, that this redef- point of view, it is crucial to look at the data inition is now made on a linguistic basis. In that was covered by earlier non-formal gram- order to achieve parsability, the surface descrip- I marians. • As far as syntactic theory is concerned, there 41n such sentence as "I asked John to go home", the is no need to reinvent the wheel. Our de- noun before the infinitive clause is analysed as the (se- ! scription has its basis in the so-called "classi- mantic) subject of the infinitive rather than as a com- plement of the governing verb. cal model" based on the work of the French SFor instance, detecting the distinct roles of the to- linguist Lucien Tesni~re. His structural model infinitive clause in the functional roles of the purpose should be capable of describing any occurring or reason is usually difficult (e.g. Quirk et al. (1985, i p. 564): "Whg did he do itf; purpose: "To relieve his 2See discussion of an earlier engineering art in apply- anger" and reason: "Because he was angry"). In such ing a dependency grammar in Kettunen (1994). sentence as UA man came to the party to have a good 3For instance, the notion of adjacency was redefined time", the interpretation of the infinitive clause depends I in WG, but was still unsuitable for "free~ word order on the interaction of the contextual and lexical semantics languages. rather than a structural distinction. I I 2 I i tion should not contain elements which can not tively adequate constitutional grammars are not be selected by using contextual information. It known. In the remaining sections, we show that is important that the redefinition should not be a descriptively adequate dependency model can made because an arbitrary mathematical model be constructed so that it is formally explicit and denies e.g. crossing dependencies between the parsable. syntactic elements. 3 Parallelism between the syntactic I 2 Constituency vs. dependency and semantic structures A central idea in American structuralism was Obviously, distributional descriptions that do to develop rigorous mechanical procedures, not contribute to their semantic analysis can i i.e. "discovery procedures", which were assumed be given to linguistic strings. Nevertheless, to decrease the grammarians' own, subjective the minimal descriptive requirement should be assessment in the induction of the grammars. that a syntactic description is compatible with I This practice was culminated in Harris (1960, the semantic structure. The question which p. 5), who claimed that "the main research arises is that if the correspondence between of descriptive linguistics, and the only rela- syntactic and semantic structures exists, why tion which will be accepted as relevant in the should these linguistic levels be separated. For I present survey, is the distribution or arrange- example, Sgall (1992, p. 278) has questioned ment within the flow of speech of some parts or the necessity of the syntactic level altogether. features relative to others." His main argument for dispensing with the I The crucial descriptive problem for a distri- whole surface syntactic level is that there are butional grammar (i.e. phrase-structure gram- no strictly synonymous syntactic constructions, mar) is the existence of non-contiguous ele- and he therefore suggests that the surface word I ments. The descriptive praxis of some earlier IC order belongs more properly to the level of mor- theoricians allows discontiguous constituents. phemics. This issue is rather complicated. We For example, already Wells (1947) discussed the agree that surface word order does not belong problem at length and defined a restriction for to syntactic structure, but for different reasons. I discontiguous constituents6. Wells' restriction In contradistinction to Sgall's claim, Mel'~uk implies that a discontiguous sequence can be (1987, p. 33) has provided some evidence where a constituent only if it appears as a contigu- the morphological marker appears either• in the 1 ous sequence in another context. This means head or the dependent element in different lan- that Wells' characterisation of a constituent de- guages, as in the Russian "kniga professor+a" fines an element which is broadly equivalent to (professor's book) and its Hungarian equivalent I the notion of bunch in Tesni~re's (1959} the- '~professzor kSnyv÷e'. Consequently, Mel'~uk ory. Consequently, these two types of grammars (1987, p. 108) distinguishes the morphological are capable of describing the equivalent syntac- dependency as a distinct type of dependency. ! tic phenomena and share the assumption that Thus morphology does not determine the syn- a syntactic structure is compatible with its se- tactic dependency, as Tesni~re (1959, Ch. 15) mantic interpretation. However, the extended also argues. constituent grammar thus no longer provides a For Tesni~re (1959, Ch. 20:17) meaning i rigorous distributional basis for a description, (Fr. sens) and structure are, in principle, inde- and its formal properties are unknown. pendent. This is backed by the intuition that We can conclude our argument by stating one recognises the existence of the linguistic I that the reason to reject constitutional gram- structures which are semantically absurd, as il- mars is that the formal properties for descrip- lustrated by the structural similarity between 6Wells (1947): "A discontinuous sequence is a con- the nonsensical sentence "Le silence vertebral i stituent il in some environment the corresponding contin- indispose la voie licite" and the meaningful sen- uous sequence occurs as a constituent in a construction tence "Le signal vert indique la voie libre". semantically harmonious with the constructions in which The independence of syntactic and seman- the given discontinuous sequence occurs." Further, Wells notes that "The phrase semantically harmonious is left tic levels is crucial for understanding Tesni~re's ill undefined, and will merely be elucidated by examples." thesis that the syntactic structure follows from m !
3 the semantic structure, but not vice versa. This The basic element in syntactic description is means that whenever there is a syntactic rela- the nucleus. It corresponds to a node in a de- tion, there is a semantic relation (e.g. comple- pendency tree. When the sentence is repre- mentation or determination) going in the op- sented as a dependency tree, the main node con- posite direction. In this view, the syntactic tains the whole verb chain. head requires semantic complementation from There are at least two reasons why the con- its dependents. Only because the syntactic and cept of the nucleus is needed. In the first place, I semantic structures belong to different levels there are no cross-linguistically valid criteria is there no interdependency or mutual depen- to determine the head in, say, a prepositional dency, though the issue is sometimes raised in phrase. One may decide, arbitrarily, that ei- the literature. ther the preposition or the noun is the head of I There is no full correspondence between the the construction. Second, because the nucleus syntactic and semantic structures because some is also the basic semantic unit, it is the minimal semantic relations are not marked in the func- unit in a lexicographical description. I tional structure. In Tesni~re (1959, p. 85), for 4.2 Linearisation example, there are anaphoric relations, seman- tic relations without correspondent syntactic re- Tesni~re makes a distinction between the linear I lations. order, which is a one-dimensional property of the physical manifestations of the language, and 4 Surface representation and the structural order, which is two-dimensional. I syntactic structure According to his conception, constructing the 4.1 The nucleus as a syntactic primitive structural description is converting the linear order into the structural order. Restricting him- The dependency syntactic models are inher- self to syntactic description, Tesni~re does not ently more "word oriented" than constituent- formalise this conversion though he gives two I structure models, which use abstract phrase cat- main principles: (1) usually dependents either egories. The notion of word, understood as an immediately follow or precede their heads (pro- orthographic unit in languages similar to En- jectivity) and when they do not, (2) additional i glish, is not the correct choice as a syntactic devices such as morphological agreement can in- primitive. However, many dependency theo- dicate the connexion. ries assume that the orthographic words directly Although Tesni~re:s distinction between the correspondr to syntactic primitives (nodes in I linear and structural order corresponds to some the trees). Although the correspondence could extent with the distinction between the linear be very close in languages like English, there are precedence (LP) and the immediate dominance, languages where the word-like units are much there is a crucial difference in emphasis with re- I longer (i.e. incorporating languages). spect to those modern syntactic theories, such TesniSre observed that because the syntactic as GPSG, that have distinct ID and LP compo- connexion implies a parallel semantic connex- nents. Tesni~re excludes word order phenom- I ion, each node has to contain a syntactic and a ena from his structural syntax and therefore semantic centre. The node element, or nucleus, does not formalise the LP component at all. is the genuine syntactic primitive. There is no Tesni~re's solution is adequate, considering that one-to-one correspondence between nuclei and I in many languages word order is considerably orthographic words, but the nucleus consists of free. This kind of "free" word order means that one or more, possibly discontiguous, words or the alternations in the word order do not neces- parts of words. The segmentation belongs to sarily change the meaning of the sentence, and I the linearisation, which obeys language-specific therefore the structural description implies sev- rules. Tesni~re (1959, Ch 23:17) argued that eral linear sequences of the words. This does the notion word, a linear unit in a speech-chain, not mean that there are no restrictions in the I does not belong to syntactic description at all. linear word order but these restrictions do not A word is nothing but a segment in the speech emerge in the structural analysis. chain (1959, Ch 10:3). In fact, Tesni~re assumes that a restriction I 7See Kunze (1975, p. 491) and Hudson (1991). that is later formalised as an adjacency princi- l I
v I ple characterizes the neutral word order when he 5.1 Gaifman's formulation says that there are no syntactic reasons for vio- The classical studies of the formal properties lating adjacency in any language, but the prin- of dependency grammar are Gaifman (1965) ciple can be violated, as he says, for stylistic and Hays (1964) 11, which demonstrate that de- reasons or to save the metric structure in poet- pendency grammar of the given type is weakly ics. If we replace the stylistic reasons with the equivalent to the class of context-free phrase more broader notion which comprises the dis- structure grammars. The formalisation of de- course functions, his analysis seems quite con- pendency grammars is given in Gaifman (1965, sistent with our view. Rather than seeing that p. 305): For each category X, there will be a fi- there are syntactic restrictions concerning word nite number of rules of the type X (YI, Y2"'" Y! * order, one should think that some languages due Fi+t-. "Yn), which means that YI"" "Y~ can de- to their rich morphology have more freedom in pend on X in this given order, where X is to using word order to express different discourse occupy the position of ,. ! functions. Thus, linearisation rules are not for- Hays, referring to Gaifman's formulation mal restrictions, but language-specific and func- above, too strongly claims that "[d]ependency tional. theory is weakly omnipotent to IC theory. The There is no need for constituents. Tesni~re's proof is due to Gaifman, and is too lengthy to theory has two mechanisms to refer to linearisa- present here. The consequence of Gaifman's tion. First, there are static functional categories theorem is that the class of sets of utterances [...] with dynamic potential to change the initial cat- is Chomsky's class of context-free languages." ! egory. Thus, it is plausible to separately define This claim was later taken as granted to ap- the combinatorial and linearisation properties of ply to any dependency grammar, and the first, each category. Second, the categories are hierar- often cited, attestation of this apparently false ! chical so that, for instance, a verb in a sentence claim appeared in Robinson (1970). She pre- governs a noun, an adverb or an adjective. The sented four axioms of the theory and claimed lexical properties, inherent to each lexical ele- they were advocated by Tesni~re and formalised ment, determine what the governing elements by Hays and Gaifman. are and what elements are governed. Thus, the over-all result of the Gaifman- There are no simple rules or principles for Hays proof was that there is a weak equiv- linearisation. Consider, for example, the treat- alence of dependency theory and context-free ! ment of adjectives in English. The basic rule is phrase-structure grammars. This weak equiva- that attributive adjectives precede their heads. lence means only that both grammars charac- However, there are notable exceptions, includ- ing the postmodified adjectives s, which follow SlTesni~re is not mentioned in these papers. Gaif- man's paper describes the results "... obtained while their heads, and some lexical exceptions 9, which the author was a consultant for the RAND Corporation usually or always are postmodifying. in the summer of 1960.~ Whereas phrase-structure sys- tems were defined by referring to Chomsky's Syntactic 5 Historical formulations Structures, the corresponding definition for the depen- dency systems reads as follows: "By dependency system In this section, the early formalisations of we mean a system, containing a finite number of rules, by the dependency grammar and their relation to which dependency analysis for certain language is done, Tesni~re's theory are discussed. The depen- as described in certain RAND publications (Hays: Febru- ary 1960; Hays and Ziehe, April 1960).~ Speaking of the dency notion was a target of extensive formal dependency theory, Hays (1960) refers to the Soviet work studies already in the first half of the 1960's l°. on machine translation using the dependency theory of Kulagina et al. In Hays (1964), the only linguistic refer- 8Example: "It is a phenomenon consistent with ..." ence is to the 1961 edition of Hjelmslev's Prolegomena: 9Example: "president el.ect" "Some of Hjelmslev's empirical principles are closely re- l°A considerable number of the earlier studies were lated to the insight behind dependency theory, but em- listed by Marcus (1967, p. 263), who also claimed that pirical dependency in his sense cannot be identified with "Tesni~re was one ol the first who used (dependency) abstract dependency in the sense of the present paper, graphs in syntax. His ideas n,ere repeated, developed and since he explicitly differentiates dependencies 1ram other precised by Y. Lecer] ~ P. l"hm (1960), L. Hirschbe~ and kinds at relations, whereas the present theorlt intends to L Lynch, particularly by studying syntactic projectivity be complete, i.e. to account lor all relations among units and linguistic subtrees." ot utterances." 'i terize the same sets of strings. Unfortunately, this formulation had little to do with TesniSre's "---- main: dependency theory, but as this result met the re- quirements of a characterisation theory, interest LIKE in the formal properties of dependency grammar diminished considerably. WOULD DO I 5.2 Linguistic hypotheses Tesni~re's Hypothesis, as Marcus (1967) calls it, assumes that each element has exactly one head. ~VHAT YOU ME TO Marcus also formulates a stronger hypothesis, I the Projectivity hypothesis, which connects the Figure 1: Non-projective dependency tree linear order of the elements of a sentence to the structural order of the sentence. The hypoth- I esis is applied in the following formulation: let Figure I violates the principle, the construction X "- ala 2 ...ai ...an be a sentence, where ai and is non-projective. aj are terms in the sentence. If the term ai is 5.3 Formal properties of a subordinate to the term aj, and there is an in- I Tesni~re-type DG dex k which holds min(i,j) < k < max(i,j), then the term ak is subordinate to the term aj. Our current work argues for a dependency This is the formal definition of projectivity, grammar that is conformant with the original I also known as adjacency or planarity. The intu- formulation in Tesni~re (1959) and contains the itive content of adjacency is that modifiers are following axioms: ! placed adjacent to their heads. The intuitive • The primitive element of syntactic descrip- content behind this comes from Behaghel's First tion is a nucleus. Law 12 (Siewierska, 1988, p. 143). The adjacency principle is applicable only if • Syntactic structure consists of connexions I the linear order of strings is concerned. How- between nuclei. ever, the target of Tesni~re's syntax is struc- • Connexion (Tesnikre, 1959, Ch. 1:11) is a tural description and, in fact, Tesnibre discusses binary functional relation between a supe- I linear order, a property attributable to strings, rior term (regent.) and inferior term (depen- only to exclude linearisation from his concep- dent). tion of syntax. This means that a formalisa- tion which characterises sets of strings can not • Each nucleus is a node in the syntactic tree I even be a partial formalisation of Tesni~re's the- and it has exactly one regent (Tesnibre, ory because his syntax is not concerned with 1959, Ch. 3:1). strings, but structures. Recently, Neuhaus and • A regent, which has zero or more depen- I BrSker (1997) have studied some formal prop- dents, represents the whole subtree. erties of dependency grammar, observing that Gaifman's conception is not compatible either • The uppermost regent is the central node I with Tesnibre's original formulation or with the of the sentence. "current" variants of DG. These axioms define a structure graph which There are several equivalent formalisations is acyclic and directed, i.e. the result is a tree. I for this intuition. In effect they say that in a These strong empirical claims restrict the the- syntactic tree, where words are printed in lin- ory. For example, multiple dependencies and all ear order, the arcs between the words must not kinds of cyclic dependencies, including mutual cross. For example, in our work, as the arc be- dependency, are excluded. In addition, there I tween the node "what" and the node "do" in can be no isolated nodes. However, it is not required that the structure l~ "The most important law is that what belongs to* gether mentally (semantically) is placed close together be projective, a property usually required in I syntactically." many formalised dependency theories that do I I 6 I not take into account the empirical fact that non-projective constructions occur in natural languages. WAS RUNNING 6 The Functional Dependency ! Grammar Our parsing system, called the Functional De- DOG IN HOUSE pendency Grammar (FDG), contains the follow- ing parts: THE THE • the lexicon, 'i Figure 2: "The dog was running in the house" • the CG-2 morphological disambiguation (Voutilainen, 1995; Tapanainen, 1996), and S • the Functional Dependency Grammar (Ta- panainen and J~rvinen, 1997; J~irvinen and ~ffll[l ill ." Tapanainen, 1997). DID RUN 6.1 On the formalism and output Ii It has been necessary to develop an expressive formalism to represent the linguistic rules that DOG IN HOUSE build up the dependency structure. The de- aep- scriptive formalism developed by Tapanainen can be used to write effective recognition gram- THE THE mars and has been used to write a comprehen- i sive parsing grammar of English. Figure 3: "Did the dog run in the house" When doing fully automatic parsing it is i necessary to address word-order phenomena. Therefore, it is necessary that the grammar for- is useful for many practical purposes. Con- |i realism be capable of referring simultaneously sider, for example, collecting arguments for a both to syntactic order and linear order. Obvi- given verb "RUN". Having the analysis such as ously, this feature is an extension of Tesni~re's those illustrated in Figure 2, it is easy to ex- theory, which does not formalise linearisation. cerpt all sentences where the governing node is Our solution, to preserve the linear order while verbal having a main element that has "run" presenting the structural order requires that as the base form, e.g. ran, "was running" (Fig- functional information is no longer coded to the ure 2), "did run" {Figure 3). The contraction canonical order of the dependentsIs. form "won't run" obtains the same analysis (the In the FDG output, the functional informa- same tree although the word nuclei can contain tion is represented explicitly using arcs with la- extra information which makes the distinction) bels of syntactic functions. Currently, some 30 as a contraction of the words "will not run". syntactic functions are applied. As the example shows, orthographic words were To obtain a closer correspondence with the segmented whenever required by the syntactic Ii semantic structure, the nucleus format corre- analysis. i sponding to Tesni~re's stemmas is applied. It This solution did not exist prior the FDG lSCompare this solution with the Prague approach, and generally is not possible in a monostratal which uses horizontal ordering as a formal device to ex- dependency description, which takes the (or- l press the topic-focus articulation at their tectogrammat- thographic) words as primitives. The problem ical level. The mapping from the tectogrammatical level is that the non-contiguous elements in a verb- to the linear order requires separate rules, called 8hallow chain are assigned into a single node while the rules (Petkevi~, 1987). Before such a description exists, one can not make predictions concerning the complexity subject in between belongs to its own node. of the grammar. For historical reasons, the representation con- tains a lexico-functional level closely similar to the syntactic analysis of the earlier English Con- straint Grammar (ENGCG) (Karlsson et al. , 1995) parsing system. The current FDG for- LOVE malism overcomes several shortcomings 14 of the earlier approaches: (1) the FDG does not rely on the detection of clause boundaries, (2) pars- JOHN MARY ing is no longer sequential, (3) ambiguity is rep- resented at the clause level rather than word level, (4) due to explicit representation of de- BILL AND AND JOAN pendency structure, there is no need to refer to phrase-like units. Because the FDG rule formal- Figure 4: Coordinated elements ism is more expressive, linguistic generalisation can be formalised in a more transparent way, which makes the rules more readable. any (governing) role in the syntactic structure as they do in many word-based forms of depen- 7 Descriptive solutions dency theory (e.g, Kunze (1975) and Mel'/:uk 7.1 Coordination (1987)). We now tackle the problem of how coordination Unlike the other arcs in the tree, the arc can be represented in the framework of depen- marking coordination does not imply a depen- dency model. For example, Hudson (1991) has dency relation but rather a functional equiva- argued that coordination is a phenomenon that lence. If we assume that the coordinated el- requires resorting to a phrase-structure model. ements have exactly the same syntactic func- Coordination should not be seen as a directed tions, the information available is similar to that functional relation, but instead as a special con- provided in Tesnibre:s representation. If needed, nexion between two functionally equal elements. we can simply print all the possible combina- The coordination connexions are called junc- tions of the coordinated elements: "Bill loves tions in Tesnibre (1959, Chs. 134-150). Tesni~re Mary", "John loves Mary ~, etc. considered junctions primarily as a mechanism 7.2 Gapping to pack multiple sentences economically into one. Unfortunately, his solution, which repre- It is claimed that gapping is even a more se- sents all coordinative connexions in stemmas, rious problem for dependency theories, a phe- nomenon which requires the presence of non- is not adequate, because due to cyclic arcs the i result is no longer a tree. terminal nodes. The treatment of gapping, :i where the main verb of a clause is missing, fol- > Our solution is to pay due respect to the for- lows from the treatment of simple coordination. mal properties of the dependency model, which requires that each element should have one and In simple coordination, the coordinator has only one head} 5 This means that coordinated an auxiliary role without any specific function elements are chained (Figure 4) using a specific in the syntactic tree. In gapping, only the coor- arc for coordination (labeled as cc). The coordi- dinator is present while the verb is missing. One nators are mostly redundant markers (Tesnibre, can think that as the coordinator represents all 1959, Ch. 39:5) 16, especially, they do not have missing elements in the clause, it inherits all properties of the missing (verbal) elements (Fig- ~4Listed in Voutilainen (1994). ure 6). This solution is also computationally lSThe treatment of coordination and gapping in Ka- effective because we do not need to postulate hane (1997) resembles ours in simple cases. However, empty nodes in the actual parsing system. this model maintains projectivity, and consequently, both multiple heads and extended nuclei, which are es- From a descriptive point of view there is no sentially phrase-level units, are used in complex cases, problem if we think that the coordinator ob- making the model broadly similar to Hudson (1991). tains syntactic properties from the nucleus that ~The redundancy is shown in the existence of asyn- detic coordination. As syntactic markers, coordinators demonstrated !)5' the existence of contrasting set of co- are not completely void of semantic content, which is ordinators; 'and', 'or', 'but' etc.
8 =!
multaneously, we explicate the formal proper- 5 three Coling-ACL '98 workshop referees for useful 9 tune and Discourse Patterns, volume 2 of <.> Theoretical and Computational Linguistics. Institute of Theoretical and Computational Figure 5: Text-based representation Linguistics, Charles University, Prague. Zellig S. Harris. 1960. Structural Linguistics. Phoenix Books. The University of Chicago it is connected to. Thus, in a sentence with ver- Press, Chicago & London, first Phoenix edi- I I tion. Formerly entitled: Methods in Struc- bal ellipsis, e.g. in the sentence "Jack painted the kitchen white and the living room blue", the tural Linguistics, 1951. coordinator obtains the subcategorisation prop- David G. Hays. 1960. Grouping and depen- erties of a verb. A corresponding graph is seen dency theories. Technical Report RM-2646, in Figure 6. The RAND Corporation, September. Due to 'flatness' of dependency model, there David G. Hays. 1964. Dependency theory: A is no problem to describe gapping where a sub- formalism and some observations. Language, ject rather than complements are involved, as 40:511-525. the Figure 5 shows. Note that gapping provides Hans Jiirgen Heringer. 1996. Empiric und Intu- clear evidence that the syntactic element is a ition bei Tesni~re. In Gertrud Greciano and nucleus rather than a word. For example, in Helmut Schumacher, editors, Lucien Tesni~ne the sentence "Jack has been lazy and Jill an- - Syntaze structurale et operations mentales, gry', the elliptic element is the verbal nucleus volume 330 of Linguistische Arbeiten, pages has been. 15-31. Niemeyer. Richard Hudson. 1989. Towards a computer- 8 Conclusion testable word grammar of English. UCL This paper argues for a descriptively adequate working papers in Linguistics, 1:321-338. syntactic theory that is based on dependency Richard Hudson. 1991. English Ilion/ Gram- rather than constituency. Tesni~re's theory mar. Basil Blackwell, Cambridge, MA. seems to provide a useful descriptive framework Timo J~rvinen and Pasi Tapanainen. 1997. A for syntactic phenomena occurring in various dependency parser for English. Technical Re- natural languages. We apply the theory and port TR-1, Department of General Linguis- develop the representation to meet the require- tics, University of Helsinki, Finland, March. ments of computerised parsing description. Si- Sylvain Kahane. 1997. Bubble trees and syn- iI 9 PAINTED 7 JACK KITCHEN WHITE AND THE LIVING_ROOM BLUE THE Figure 6: Jack painted the kitchen white and the living room blue. tactic representations. In Becker and Krieger, Leech, and Jan Sx~rtvik. 1985. A Compre- editors, Proceedings 5th Meeting of Mathe- hensive Grammar of the English Language. matics of language, Saarbriicken, DFKI. Longman, Harcourt. Fred Karlsson, Atro Voutilainen, Juha Heikkil~i, Jane J. Robinson. 1970. Dependency struc- and Arto Anttila, editors. 1995. Constraint tures and transformational rules. Language, Grammar: a language-independent system 46:259-285. for parsing unrestricted text, volume 4 of Petr Sgall. 1992. Underlying structure of sen- Natural Language Processing. _Mouton de tences and its relations to semantics. Wiener Gruyter, Berlin and New York. Slawistischer Almanach, Sonderband 30:349- Kimmo Kettunen. 1994. Evaluating FUNDPL, 368. a dependency parsing formalism for Finnish. Anna Siewierska. 1988. Word Order Rules. I- In Research in Humanities Computing, vol- Croom Hehn, London. ume 2, pages 47-63. Clarendon Press, Oxford. Pasi Tapanainen and Timo J~irvinen. 1997. A Jiirgen Kunze. 1975. Abh6ngigkeitsgrammatik. non-projective dependency parser. In Pro- Akadenaie-Verlag, Berlin. ceedings of the 5th Conference on Applied Solomon Marcus. 1967. Introduction Natural Language Processing, Washington, mathdmatique h la linguistique structurale. D.C, pages 64-71, Washington, D.C.,April. Dunod, Paris. Association for Computational Linguistics. James D. McCawley. 1982. Parentheticals and Pasi Tapanainen. 1996. The constraint gram- discontinuous constituent structure. Linguis- mar parser CG-2. Publications 27, Depart- tic Inquiry, 13(1):91-106. ment of General Linguistics, University of Igor A. Mel'~.uk. 1987. Dependency Syntax: Helsinki, Finland. Theory and Practice. State University of New Lucien Tesni~re. 1959. Elgments de syntaxe York Press, Albany. structurale. Editions Klincksieck, Paris. Peter Neuhaus and Norbert BrSker. 1997. The Atro Voutilainen. 1994. Designing a Parsing complexity of recognition of linguistically Grammar. Publications of Department of adequate dependency grammars. In ACL- General Linguistics, University of Helsinki, EACL'97 Proceedings, pages pp. 337-343, No. 22, Helsinki. Madrid, Spain, July. Association for Compu- Atro Voutilainen. 1995. Morphological disam- tational Linguistics. biguation. In Karlsson et al. (1995), chap- i Vladimir Petkevi~:. 1987. A new dependency ter 6, pages 165-284. based specification. Theoretical Linguistics, Rulon S. Wells. 1947. Immediate constituents. 14:143-172. Language. Reprinted in Martin Joos: Read- Randolph Quirk, Sidney Greenbaum, Geoffrey ings In Linguistics I, 1957, pp. 186-207. 10 i