{ "paper_id": "W98-0138", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:04:29.468788Z" }, "title": "Comparison of XTAG and LEXSYS Grammars", "authors": [ { "first": "Martine", "middle": [], "last": "Smets", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Sussex", "location": { "postCode": "BNl 9HQ", "settlement": "Brighton", "country": "UK" } }, "email": "martines@cogs.susx.ac.uk" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "W98-0138", "_pdf_hash": "", "abstract": [], "body_text": [ { "text": "This paper presents work that forms part of the ongoing LEXSYS project on wide-coverage parsing, 1 and more precisely, some differences between our D-Tree grammar and XTAG 1995.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We use the Lexicalised D-Tree Grammar (LDTG) formalism (Rambow et al. 95) , which is based on the Lexicalized Tree Adjoining Grammar (LTAG) formalism. In LDTG, there are two types of edges between nodes: d-edges, represented with a broken line, and p-edges, represented by a solid line. Trees are combined by two substitution-like operations, both of which involve combining two descriptions, by equating exactly one node from each description. One of the operations is always used to add complements, and involves equating a frontier node (in the d-tree that is getting the complement) with the root of some component (in the d-tree that is providing the complement ), such that the two nodes being equated are compatible . An example of substitution is shown in Figure 1 .", "cite_spans": [ { "start": 55, "end": 73, "text": "(Rambow et al. 95)", "ref_id": null } ], "ref_spans": [ { "start": 764, "end": 772, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Grammar Formalism", "sec_num": "2" }, { "text": "The d-tree for to adore is composed with the d-tree for seems by equating the two nodes labelled VP [fin: -] . The top component of the to adore tree can then be fitted into the resulting d-tree by equating the root of the seems tree with the lower S of the to adore tree.", "cite_spans": [ { "start": 100, "end": 108, "text": "[fin: -]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Grammar Formalism", "sec_num": "2" }, { "text": "A second operation is used to add modifiers, but we are not going to discuss it in this paper.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grammar Formalism", "sec_num": "2" }, { "text": "3 Differences between XTAG and LEXSYS Grammars", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Grammar Formalism", "sec_num": "2" }, { "text": "A first difference between our DTG and TAG is that we do not clairn that elementary trees express in all cases the predicate-argument structure of their anchor; instead, they represent the syntactic requirements of their anchor. To illustrate, because raising verbs subcategorize for a syntactic subject, they anclior a standard verb tree with a subject, and not a tree rooted in VP without a subject, as in TAG. On the other hand, there are trees rooted in VP which represent VP complements and can be anchored by any verb. In those trees, there is no subject (because VP complements do not have syntactic subjects), and a semantic argument of the verb is thus rnissing. This choice allows us to adopt other linguistic analyses than the ones supported by XTAG, as will be shown in the next sections.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Trees Are Syntactic Representations", "sec_num": "3.1" }, { "text": "Distance Dependency A main difference between the two grammars is that there are VP complements in our grammar, when there are only S complements in XTAG ( except for auxiliaries and raising verbs). To the sentence in (1), our grammar gives the analysis in (la), while XTAG gives the analysis in (lb). such as Head-Driven Phrase Structure Grammar (HPSG, Pollard and Sag 1994) and Lexical Functional Grammar (LFG, Bresnan 1982) whiie the analysis in (lb) is the analysis of Government and Binding (GB, Haegeman 1991) . Arguments given in the XTAG report for the representation in ( 1 b) include a uniform treatment for indicative, infinitive and gerund embedded clauses (XTAG report 1995 (XTAG report , 1998 . This implies that both infinitive and gerunds are analyzed as having an empty subject, which is questionable, because there is no evidence for the existence of PR0 2 ; it is even more questionable for gerunds, which have the same distribution as NPs (this is true even for verbal gerunds), and can hardly be characterized as clauses (Malouf 1997 inter alia ).", "cite_spans": [ { "start": 354, "end": 375, "text": "Pollard and Sag 1994)", "ref_id": null }, { "start": 380, "end": 426, "text": "Lexical Functional Grammar (LFG, Bresnan 1982)", "ref_id": null }, { "start": 501, "end": 515, "text": "Haegeman 1991)", "ref_id": "BIBREF2" }, { "start": 669, "end": 686, "text": "(XTAG report 1995", "ref_id": null }, { "start": 687, "end": 706, "text": "(XTAG report , 1998", "ref_id": null }, { "start": 1042, "end": 1054, "text": "(Malouf 1997", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": ": V V P : ( f i v , n \u2022 --] 1 / VP:[fon:-) sums /'-.... v NPi 1 1 hotdogs s /'-..... ~p VP YOY /'-..... V S 1 want ~ NPi S 1 ~ wh\u2022I NP VP 1 /'-..... PRO V VP 1 / \" -. V", "eq_num": "NPi" } ], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "An important reason for XTAG to adopt the analysis in (1 b) is that it seems to be the only type of analysis possible in that formalism (except if equi verbs like want anchor the elementary tree for raising verbs ). This comes from the fact that unbounded dependencies which extend across more than one clause boundary are achieved through the use of auxiliary trees in XTAG: to derive the sentence in (2), an initial tree for buy is combined with an auxiliary tree for want (Figure 1 ).", "cite_spans": [], "ref_spans": [ { "start": 475, "end": 484, "text": "(Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "(2) What do you yant to buy?", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "2 PRO, besides being unmotivated, creates theoryinternal problems: XTAG has to define two different infinitive auxiliaries to, one tvhich assigns the case no case (when the subject is PRO) and the other one which does not assign a.ny case (when the complementizer /or assigns accusative case .to the subject). This distinction between two to is of course ad-hoc .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "The auxiliary tree for want is grafted onto the lower S of the buy tree, and the recursivity of the process creates unbounded dependency. And because in auxiliary trees the root node and the foot node must be of the same category, verbs such as want cannot take a VP complement ( assuming want anchors an S-tree).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "In our grammar, on the other hand, there is no such restriction, and verbs can take S complements as weil as VP complements. This decision to introduce VP complements in the grammar has a number of consequences ( some of which are related to what was discussed m section 1 ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "\u2022 auxiliaries and ra1smg verbs anchor the same tree farnily as other verbs which take VP complements;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "\u2022 passive trees are rooted in VP;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "\u2022 because trees for auxiliaries and ra1smg verbs are rooted in S as any other verb tree, there are no predicative trees;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "\u2022 the grammar has at least twice as many verb trees as XTAG 95 ( each tree rooted in S has a counterpart rooted in VP), and in fact, more than that as we use multiple instances of the same tree to represent disjunctive feature values. \u2022 Each of tliese points will be addressed in the next sections.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Complementation and Long", "sec_num": "3.1.1" }, { "text": "Another type of construction for which we assume the existence of a VP complement is thc Subject-to-Object Raising (SOR) structure illustrated in (3).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "(3) We believe [Kirn] [to be very smart]", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "In that analysis 3 , which we adopt, raising verbs such as believe have two cornplements, an NP and a VP. In the XTAG analysis, SOR verbs have only one cornplernent, a clause. We assign the sarne kind of analysis to another type of verbs, referred to as verbs of considering in Pollard and Sag ( 1994) : consider in (4) and regard in (5) have two complernents, an NP and respectively an AP and a PP.", "cite_spans": [ { "start": 278, "end": 301, "text": "Pollard and Sag ( 1994)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "(4) I consider Jack quite intelligent (5) We regard him as a nuisance This analysis has been debated since the early seventies, and supported by a number of researchers, Pollard and Sag (1994) arnong others.", "cite_spans": [ { "start": 170, "end": 192, "text": "Pollard and Sag (1994)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "XTAG, on the other hand, adopt the GB analysis, which considers that verbs of considering and the like have only one cornplernent, a srnall clause. Srnall clauses are Ss headed by an ernpty verb, and anchored by the complement of that verb (NP, PP or AP). This account is not without problems. First, it has to postulate a.n unmotivated ernpty verb position: there is no evidence that such a position should exist. Its purpose is to allow adjunction of raising and auxiliary verbs, but this is a purely technical device which is not supported by linguistic evidence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "A more irnportant problem is the fact that verbs which take small clause complements rnust be able to constrain the srnall clause predicate: consider allows PPs,. NPs and APs (6) while prefer allows PPs only (7). 6We consider Kirn a good teacher We consider Kirn quite good We consider Kirn out of his rnind 7*We pref er Kirn a good teacher *We prefer Kirn quite good We prefer Kim out of here Verbs who subcategorize for clausal cornplernents cannot specify the subcategorization requirements of the verb in the complement clause; for exarnple, there is no exarnple of a verb like say which would stipulate what kind of complernent the verb in its clausal complernent should have. Accordingly, in the XTA G account, the clausal cornplement is not expanded, whether it is a standard clause or a srnall clause. But the data in (6) and 7show that verbs of considering and the like do select the type of phrases which follow the NP; the solution adopted in XTAG is to use the feature mode ( whose values are usually indicative, imperative, subjunctive, etc.) and to add to the range of features nom and prep (for NP and AP, and PP respectively). The verb consider selects an S which has a feature mode with value nom/prep, while prefer selects a small clause with prep as value for the feature rnode. Of course, the decision to add these values to the range of values of the feature rnode is ad-hoc, as they have nothing to do with verb rnode, and are only a technical device to match the subcategorization requirernents of the verb of considering with the actual category of the cornplernent in the embedded srnall clause. Our solution, on the other hand, is straightforward: if the verb consider constrains the type of phrase that follows the NP it is because this phrase is also one of its complernents.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "Our choice of analysis, besides being straightforward and rnotivated by the data, also allows for a rnore uniform account of passive: the passive of verbs of considering and the like is handled by the same lexical rules as for other transitive verbs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Verbs of Considering", "sec_num": "3.2" }, { "text": "In XTAG, raising verbs and auxiliary verbs anchor the sarne auxiliary tree rooted in VP. In our grarnrnar, on the other hand, those verbs anchor trees rooted in S, and belong to different farnilies.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Auxiliaries and Raising Verbs", "sec_num": "3.3" }, { "text": "There have been debates in the literature about the status of auxiliary verbs, and. several authors have argued that auxiliaries and rnodals should be considered as rnain verbs (Pullum and um 8~~ 11077' f':!\"..,...ia\" \"~ al r1 OQ'l))", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Auxiliaries and Raising Verbs", "sec_num": "3.3" }, { "text": "A \"rr\"_ rnents include the fact that sorne auxi\u00fcaries behave also like main verbs ( be and have, ought, is in is to ), and the existence of serni-auxiliaries ( need, used, dare and have to 4 ) which behave like main verbs in certain environments and like auxiliaries in other environments. So, the distinction between auxiliaries and main verbs is not clear-cut, and either the tree family for auxiliary verbs will include verbs which do not always behave like auxiliaries, or verbs classified as main verbs will share characteristics with auxiliary verbs. In both cases, the obvious solution is to abandon the distinction between main verbs and auxiliaries in terms of dra.stically different types of tree, and adopt instead a unified representation for both kinds of verbs.", "cite_spans": [ { "start": 158, "end": 192, "text": "( need, used, dare and have to 4 )", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Auxiliaries and Raising Verbs", "sec_num": "3.3" }, { "text": "A second issue is the fact that in the tree for auxiliaries and raising verbs, the complement of the anchor is a VP. This implies that all subject raising verbs subcategorize for VP, which is clearly not the case ( become subcategorizes for AP or NP, turn out for AP, NP or VP). Thus, in order to get the right distribution of subcategorization, constraints on the complement of the raising verbs have to be expressed through percolation of the mode feature, which use has already been shown to be ad-hoc in similar instances.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Auxiliaries and Raising Verbs", "sec_num": "3.3" }, { "text": "There are no predicative trees in our grammar: this is a consequence of our decision to adopt a tree rooted in S for both raising verbs and auxiliaries. Also, we want a uniform treatment of predicative complernents, and this would not be the case if we adopted different trees for predicative complements of verbs of considering and predicative complernents of other types of verbs. So, predicative complements just substitute in the tree of their governing verb, like other types of complements.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicative Trees", "sec_num": "3.3.1" }, { "text": "A main criticism of our approach will be that the basic trees do not express all semantic relations: a predicative complement places semantic restrictions on the subject, and this cannot be captured in the basic trees, because predicative cornplements are substituted in the tree for the auxiliary /raising verb; similarly, for the VP complement trees, which do not have a subject 5 ; finally, in the case of passive, the passive participle anchors a VP tree too, and the subject is not e?'pressed either in the elementary tree.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicative Trees", "sec_num": "3.3.1" }, { "text": "We agree with this, but we do not claim that we can express every type of relation between constituents in basic trees; instead, we believe that it is impossible to capture all relevant information, syntactic and semantic, in the basic trees. We therefore adopt a modular representation, with the basic trees expressing mainly syntactic information, and the derivation tree most of the semantic information. We hope that this division of labour will allow US to express motivated syntactic analyses in the grammar, without having to compromise in order to also express at the same level semantic relations.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Predicative Trees", "sec_num": "3.3.1" }, { "text": "This paper has presented some differences between XTAG and the grammar we are developing in the LEXSYS project. lt has shown that the DTG formalism gives us the possibility to adopt linguistic analyses which have proven to be more motivated than the GB ones (which can also be expressed with the same formalism).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "3.4" }, { "text": "The fact that we will have much more trees than TAGs might seem like a drawback to our approach. But Evans and Weir (1998) are explorlng ways to allow a compact representation of the grammar for parsing purposes.", "cite_spans": [ { "start": 101, "end": 122, "text": "Evans and Weir (1998)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "3.4" }, { "text": "This work is supporte.d by UK EPSRC project GR/K97400 'Analysis of Naturally-occurring English Te:d with Stochaslic Lexicalized Grammars' {http://vvv.cogs.susx.ac.uk/lab/nlp/dtg/).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The SOR analysis has been advocated with compelling arguments byBresnan (1982),Postal and Pullurn (1988) andPollard and Sag (1994) inter alia.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Actually, have to behaves like a main verb in all t:nviionrnents, but has a mcaning very similar to musl, This shows that wl\u00fcch verbs are auxiliaries cannot be predicted from semantic information alone, as was noted byPullum and Wilson (1977).~r do not see any advantage of having PRO instead", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Control and cornplernentation", "authors": [ { "first": "Joan", "middle": [], "last": "Bresnan", "suffix": "" } ], "year": 1982, "venue": "The Mental Representation of Grammatical Relations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bresnan, Joan. 1982. Control and cornplernen- tation. In J. Bresnan, eds., The Mental Rep- resentation of Grammatical Relations.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "A structure-sharing parser for lexicalized grammars", "authors": [ { "first": "Roger", "middle": [], "last": "Evans", "suffix": "" }, { "first": "David", "middle": [], "last": "Weir", "suffix": "" } ], "year": 1998, "venue": "COLING/ACL'98", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Evans, Roger and David Weir. 1998. A structure-sharing parser for lexicalized gram- mars. In COLING/ACL'98.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "lntroduction to Government and Einding Theory", "authors": [ { "first": "Liliane", "middle": [], "last": "Haegeman", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Haegeman, Liliane. 1991. lntroduction to Gov- ernment and Einding Theory. Blackwell Pub- lishers, Oxford, UK.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "A constructional approach to English verbal gerunds", "authors": [ { "first": "Rob", "middle": [], "last": "Malouf", "suffix": "" } ], "year": 1996, "venue": "Proceedings of the Twenty-second Meeting of the Berkeley Linguistics Society", "volume": "", "issue": "", "pages": "255--266", "other_ids": {}, "num": null, "urls": [], "raw_text": "Malouf, Rob. 1996. A constructional approach to English verbal gerunds. In Proceedings of the Twenty-second Meeting of the Berkeley Linguistics Society, 1996, pages 255-266.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "hrasc Structure Grammar. Chicago U niversity Press", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "hrasc Structure Grammar. Chicago U niver- sity Press, Chicago, IL. of no subject, though.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Expletive noun phrases in subcategorized positions", "authors": [ { "first": "Geoffrey", "middle": [], "last": "Pullum", "suffix": "" }, { "first": "Paul", "middle": [], "last": "Postal", "suffix": "" } ], "year": 1988, "venue": "Linguistic Inquiry", "volume": "19", "issue": "", "pages": "635--670", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pullum, Geoffrey and Paul Postal. 1988. Exple- tive noun phrases in subcategorized positions. In Linguistic Inquiry, 19, pages 635-670.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Autonomous syntax and the analysis of auxiliaires", "authors": [ { "first": "Geoffrey", "middle": [], "last": "Pullum", "suffix": "" }, { "first": "Deirdre", "middle": [], "last": "Wilson", "suffix": "" } ], "year": 1977, "venue": "Language, v.53,4", "volume": "", "issue": "", "pages": "741--788", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pullum, Geoffrey and Deirdre Wilson. 1977. Autonomous syntax and the analysis of aux- iliaires. In Language, v.53,4, pages 741-788.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Parsing D-Tree Grammars", "authors": [ { "first": "Owen", "middle": [], "last": "Rambow", "suffix": "" }, { "first": "K", "middle": [], "last": "Vijay-Shanker", "suffix": "" }, { "first": "David", "middle": [], "last": "Weir", "suffix": "" } ], "year": 1995, "venue": "Proceedings of the 33rd Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rambow, Owen, K. Vijay-Shanker and David Weir. 1995. Parsing D-Tree Grammars. In Proceedings of the 33rd Meeting of the Asso- ciation for Computational Linguistics (ACL", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "num": null, "text": "Example of unbounded dependency in DTG (left) and in TAG (right)", "uris": null } } } }