{ "paper_id": "W98-0107", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:04:23.772479Z" }, "title": "Defining DTG derivations to get semantic graphs", "authors": [ { "first": "Marie-Hel~ne", "middle": [], "last": "Candito", "suffix": "", "affiliation": { "laboratory": "", "institution": "Universite Paris 7", "location": { "addrLine": "2, place Jussieu", "postCode": "7003, 75251, Cedex 05", "settlement": "Paris" } }, "email": "" }, { "first": "Sylvain", "middle": [], "last": "Kahane", "suffix": "", "affiliation": { "laboratory": "", "institution": "Universite Paris 7", "location": { "addrLine": "2, place Jussieu", "postCode": "7003, 75251, Cedex 05", "settlement": "Paris" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "", "pdf_parse": { "paper_id": "W98-0107", "_pdf_hash": "", "abstract": [], "body_text": [ { "text": "Tue aim of this paper is to find a fonnalism of the TAG family, where the derivation controller can be interpreted as a semantic dependency graph, in the sense of Meaning-Text Theory (ZM67; M88) . In a previous paper (CK98), we study tliis interpretation of the derivation tree (DT) in the case of standard TAG. We prove that, in the general case, if the predieate-argument cooccurence principle 1 [= P ACP] holds and if elementary trees correspond to a semantic unit (A91) , substitution arcs can be read as \u2022 semantic dependencies where the dependent is the anchor of the substituted tree, and adjunction arcsof any type-can be read as semantic dependencies in the opposite direction. Yet we also characterized cases where the DT shows wrong (semantic) dependencies (cf also (RVW95)). A problem may occur when, in the same sentence, clausal complementation is handled both with substitution of an embedded clause and with adjunction of a main verb.2 Further, there are well-known cases that TAG cannot handle if the PACP holds (e.g. clitic climbing in Romance (B98) , Kashmiri wh-extraction (RVW95), extraction out of NP in French (A98) . Finally, in some cases, the argumental positions in a tree are not filled by the right arguments, and thus the derivation tree does not show the right semantic dependencies (pied-piping (CK98)).", "cite_spans": [ { "start": 183, "end": 189, "text": "(ZM67;", "ref_id": "BIBREF10" }, { "start": 190, "end": 194, "text": "M88)", "ref_id": null }, { "start": 468, "end": 473, "text": "(A91)", "ref_id": null }, { "start": 1062, "end": 1067, "text": "(B98)", "ref_id": null }, { "start": 1133, "end": 1138, "text": "(A98)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": null }, { "text": "(RVW95) have defined D-tree Grammars (DTO) by ruling out predicative adjunction (e.g. adjunction of bridge verbs). Thus, DTO seems a good candidate for our goal. 3 In Section l, we recall DTO operations and 1 A tree anchored by a predicate must contain positions for all and only its arguments. l For a sentence such as That Paul wanted to stay surprised Mary, the DT shows the wrang dependencies if the tree for surprise has a substitution node for its subject., and the one for want has a foot node for its embedded clause (CK98) . Another problem occurs wilh a raising verb that serves as semantic argument to a bridge verb as in Pa11i ciaims Mary seems to cuiore hotdogs (adapted from (RVW95)). To get the correct semantic dependencies, the trees for claims and seems should combine together (either via substitution of seems or adjunction of claims) but this is impossible in TAG since seems is represented by a VP-rooted tree. study the case of relative clause interacting with n bridge verb, which has proved to be correct!y handlcll by TAG, as far as semantic dependencies arc concemed (CK98) . This leads us to propose nn extension of DTO, called GAG for Graph-driven Adjunction Grammar, whose derivation controllers are graphs (Section 2). Finally, in Section 3. we develop an original analysis of wh-words in GAG.~", "cite_spans": [ { "start": 162, "end": 163, "text": "3", "ref_id": null }, { "start": 525, "end": 531, "text": "(CK98)", "ref_id": "BIBREF5" }, { "start": 1094, "end": 1100, "text": "(CK98)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": null }, { "text": "DTO (RVW95) handles both clausal and nominal complementation with the same operation. a generalized substitution, called subsertion. and thus avoids the use of predicative adjunction. In order lo cover the long-distance dependency data (including cases not handled in TAG), this operation allows pieces of the substituted element to eome in between elements of the tree receiving substitution. induce the same non-oriented graph. we will stully a case of mismatch in Section l and 3. 4 We are thankful to Owen Rambow and David Wcir for valuable discussions about this work.", "cite_spans": [ { "start": 484, "end": 485, "text": "4", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Generalized substitution and generalized adjunction", "sec_num": "1." }, { "text": "s NJ, sJ ~ . A~- N0s ; J, si ==> s s ,,,.. ~ /\"\"-. lhink NJ, y ~! NJ.-. VI think S write ~ NJ.- V", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Generalized substitution and generalized adjunction", "sec_num": "1." }, { "text": "substituted node or placed above the root node. Fig. l shows an example of subsertion. 5 Now, ruling out predicative adjunctions implies to reconsider cases that were correctly handled by TAG as far as semantic dependencies are concemed. For example, in order to handle extraction out of a modifier (e.g. preposition stranding) and \u00ab extraction of a modifier \u00bb, we define a parallel generalization of the adjunction operation. 6 We will thus refer to generalized substitution and generalized adjunction. Fig. 2 shows the generalized adjunction of in [lhis bed} for the sentence: 7", "cite_spans": [ { "start": 427, "end": 428, "text": "6", "ref_id": null } ], "ref_spans": [ { "start": 48, "end": 54, "text": "Fig. l", "ref_id": null }, { "start": 504, "end": 510, "text": "Fig. 2", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Generalized substitution and generalized adjunction", "sec_num": "1." }, { "text": "( 1) In this bed, 1 think I have slept twice.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Generalized substitution and generalized adjunction", "sec_num": "1." }, { "text": "To get the semantic dependency between i11 and slepl, we want \u00dfin to adjoin in o:slept, still allowing a piece of the modifier (here the whole modifier) to be inserted higher. (2) I bought the books which Peter thi11ks Mary wrote.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Generalized substitution and generalized adjunction", "sec_num": "1." }, { "text": "A \u00c4 fj ,r-S \u00c4 f P N.l-S~0v => p N.l-S, Jn . sllep ln N{'\\I \u00dfm cxsleep (extraposed) sleep", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Al-", "sec_num": null }, { "text": "In TAG, the relation between a verband a relativized complement is localized. So for instance to handle (2), wrote anchors an NP modifier tree (thus an auxiliary tree) in which the bridge verb thinks adjoins (K87) . In DTO, bridge verbs receive their clausal complement via substitution. Thus in order to keep the semantic dependency between a verb and a relativized complement, we propose to allow a d-tree to substitute in a d-tree and adjoin in another one. 8 We will call this extension GAG.", "cite_spans": [ { "start": 208, "end": 213, "text": "(K87)", "ref_id": null }, { "start": 461, "end": 462, "text": "8", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Al-", "sec_num": null }, { "text": "5 Figure 1 shows d-trees that are inspired from the d-trees proposed by (RVW95) to handle a sentence such as Childre11's books Peter thinks Mary wrote. For sake of simplicity VP nodes are omitted. 6 In (RVW95), modifiers are handled by sisteradjunction, an operation that is equivalent to adding at a 2iven node a !eft-most or right-most daughie1 node. -We prefer to maintain adjunction (as in TAG), notably because we want to be able to adjoin the tree for glass-of for instance (CK98). 7 In this example, the bottom component of \u00dfin is reduced to the foot node, which is also the rope (see definition in Section 2). ' As a referee pointed out to us, there is nn a!ternate derivation of (3) in DTO in which thi11ks is adjoined", "cite_spans": [ { "start": 488, "end": 489, "text": "7", "ref_id": null } ], "ref_spans": [ { "start": 2, "end": 10, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Al-", "sec_num": null }, { "text": "GAG is an extension of DTO that uses \u2022!ir-same elementary structures, namely d-trees. But in GAG. some nodes of an elementary d-tree must be marked as being ropes (underlined in the figures). Only roots of components can be ropes. Any component containing a foot node has a root which is a rope. A component without foot node is subsitutab!e if and only if its root is a rope. In (RVW95), all the components of an elementary d-tree are considered substitutable, namely each component's root is a rope, but a d-tree can be subserted only once. namely all ropes are mutually exclusive. In our extension. d-trees with n mutually exclusive ropes are expandec.l in n d-trees with a single rope. Further. a d-tree may have several ropes which are not mutually exclusive. that is, that can each be combined wilh a separate d-tree. To sum up, GAG is a multi-rope DTG. From the Iinguistic point of view, we foresee the use of one-rope and two-rope d-trees only. Examples of two-rope d-trees will be given in Section 3. Let us now define the derivation graph (00). which is a structure that partially encodes a GAG deri \\'ation (and that we will interpret as a semantic graph).~ 1f a two-rope d-tree substitutes in a one-rope d-tree. we obtain a two rope derived d-tree and nothing in the DG tells us from which elementary d-tree euch rope comes from. Thus in GAG, the odginal elementary d-tree for each node of a derived d-tree is memorized. We thus have to specify what happens in the case of node unification during substitution or adjunction. In case of substitution, a rope unifies with n substitution site. We then consider that the resulting node comcs from the elementary d-tree that is substituted. In case of generalized adjunction. the node rece1vmg adjunction is replaced by a component of ehe adjoined tree. In the derived tree, we consider that the root of that adjoined component belongs to the tree receiving adjunction. Tue DG can now be defined ns follows: !et y be an elementary d-tree. Let q> be a derived tree. and \\jf the corresponding derivation graph (DG). Ir qi substitutes (resp. adjoins) in y, one of its ropes is used up. Let a be the name of the d-tree from which this rope originales. Tue resulling DG \\lf' is the DG '!I plus n to book (creating the syntactic attachment} anc.l wrote subserted into thinks with the relative pronoun being inserted into the right place and recciving coreference with books through features (thus neating the semantic attachment). ~ The equivalent in DTG is called a SA-tree. In GAG. it is a graph due to multi-rope d-trees. substitution (resp. adjunction) arc between ex. and y (y being the mother node). 1\u00b0 Consequently to this definition, a d-tree has as many mother nodes in the final DG as it has used ropes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "GAG: a multi-rope DTG", "sec_num": "2." }, { "text": "As in DTO, the derivation succeeds if the d-edges of the derived tree can be collapsed (forgetting the fact that some nodes can be rope nodes). From the computational point of view it can be noted that the ropes of a multi-rope d-tree can combine in whatever order with other d-trees.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "GAG: a multi-rope DTG", "sec_num": "2." }, { "text": "As we said, the main motivation for GAG is to have a formalism inspired by TAG whose derivation controllers induces semantic dependency graphs. In order to achieve that, we have relaxed the constraint that these controllers be trees.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "As linguistic constraints for elementary structures, in addition to the PACP, we type the argumental. positions as foot nodes and substitution nodes on purely linguistic grounds. 11 Generalized substitution is used for elements that are subcategorized and to which a thematic role is asslgned, while generalized adjunction is used for modifiers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "In the following, we concentrate on examples of GAG analysis involving wh-words. Consider: In these four examples we have a clause of the fonn Peter thinks Mary . wrote [book] . The distribution of this clause depends on the extracted element: for example, in (3b) that clause is an NP modifier because of the relative wh-word which and in (3c) that clause can be the syntactic argument of wonder because of the interrogative wh-word which. The (T59) analysis of relative and (indirect) interrogative clauses is that the wh-word plays two rotes: on one hand, it fills a position in the clause as pronoun and on the other hand it controls the distribution of the clause and is thus its syntactic head.", "cite_spans": [ { "start": 169, "end": 175, "text": "[book]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "We claim that it is possible (though not mandatory) to have an analysis where the particular distribution of wh-clauses is completely assumed by the wh-word.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "To do this, we represent wh-words with two-rope eiementary d-trees: the first rope will be linked to the lD lt can be noted that because we remember the origin of each node of a derived tree, a derivation need not be bottom-up.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "main clause and the second one will be linked to the phrase showing extraction. So for a relative wh-word. the first rope is a foot node which adjoins on the antecedenr iind the second rope substitutes or adjoin:,; in the phrase showing extraction, depending on thc complementlmodifier nature of the extracted element.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "For an interrogative wh-word, the first rope substitute:,; in the verb which subcategorizes for the interrogative clause. We give example of relative clauses only:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "(4a) l know the books which Mary wrote. (4b) l know the bed in which Peter slept. (4c) l know the books whose authors are famous. (4d) l know the man whose car Peter borrowed. (4e) l know the place where Peter was born. Fig. 3 shows the two-rope d-trees for the wh-words involved (the ropes are underlined).", "cite_spans": [], "ref_spans": [ { "start": 220, "end": 226, "text": "Fig. 3", "ref_id": null } ], "eq_spans": [], "section": "Taking advantage of GAG to analyse extraction", "sec_num": "3." }, { "text": "/\"\"-.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A N", "sec_num": null }, { "text": "/\"\"-. /\"\"-. The analysis for (3b), (4a) and (4b) use the same d-tree for which, \u00dfcx.which, which substitutes respectively in the d-trees for wrole and in (Fig. 4} . ", "cite_spans": [], "ref_spans": [ { "start": 154, "end": 162, "text": "(Fig. 4}", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "A N", "sec_num": null }, { "text": "N\"' ~ N\"' s N* ~ N* ~ ' 1 1 1 ri ll N s A 1 1 A which whose J anl+ N* lpanr+ S*", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A N", "sec_num": null }, { "text": "Building on DTG and TAG, we havc defincd a fonnalism, GAG, where the derivation controller can be seen as a semantic dependency graph, with thc reading defined in (CK98) . This allows us to proposc an analysis in which the distribution of clauses containing wh-words is totally controlled by the d-trees associated with the wh-words themselves. Thus topicaiization, relativization, (direct or indirecl) interrogation and cleft clauses can be handled with the same elementary d\u2022trees for verbs. Computational properties of GAG need a further study.", "cite_spans": [ { "start": 163, "end": 169, "text": "(CK98)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": null }, { "text": "In (RVW95), one motivation was to get (deep) syntactic dependencies. Though in most cases semantic and deep syntactic dependency structures", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "This is possible because we allow the derivalion controller to be a graph and use the generalized substitution.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We advocate that a determiner which is not an argument is adjoined. lt is the case for the possessive when it refers to a possessor.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This last operation arises typically for some relative pronouns. In the \u00dfo: d-trees of Fig. 3 , the foot node does not represent a semantic argument of the anchor, but a duplication of the anchor itself (the antecedent in syntax). Thus tne correspondant adjunction arc in the derivation controller (eg. the adjunction arc in Fig. 5 ) has to be collapsed in order to get the semantic graph ( Fig. 6 ).\u00dfa whose a Mary Figure 5 : GAG derivation graph l k11ow the books which Peter thinks Mary wroteFigure 6: MTT semantic graph", "cite_spans": [], "ref_spans": [ { "start": 87, "end": 93, "text": "Fig. 3", "ref_id": null }, { "start": 325, "end": 331, "text": "Fig. 5", "ref_id": null }, { "start": 391, "end": 397, "text": "Fig. 6", "ref_id": null }, { "start": 416, "end": 424, "text": "Figure 5", "ref_id": null } ], "eq_spans": [], "section": "annex", "sec_num": null }, { "text": "In 4c, whose is an argument of authors, thus \u00dfo:whose substitutes in the autlwrs tree. In (4d), we consider that whose is a lexicalization of the twoplace Semanteme 'own'. lts d-tree \u00df\u00dfwhose 12 adjoins twice, on the trees for both its arguments (here lexicalized by man and car). ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "I know the books which Peter thinks Mary wrote", "sec_num": null }, { "text": "To get the semantic graph, the two adjunction arcs of \u00df\u00dfwhose are interpreted as semantic dependencies from the adjoined tree to the tree receiving adjunction (Fig. 8) . Similarly, \u00df\u00dfwhere corresponds to a semanteme 'location' (= 'is located in') with two arguments.In the anaysis we have shown, features must be added to control which components can be inserted in a d-edge (cf. the subsertion insertion constraints in", "cite_spans": [], "ref_spans": [ { "start": 159, "end": 167, "text": "(Fig. 8)", "ref_id": null } ], "eq_spans": [], "section": "I k11ow the man whose car Peter borrowed", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Une LTAG pour le franr;ais. Ph.D. thesis. Univ. Paris", "authors": [ { "first": "A", "middle": [], "last": "Abeille", "suffix": "" } ], "year": 1991, "venue": "", "volume": "7", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Abeille, 1991 : Une LTAG pour le franr;ais. Ph.D. thesis. Univ. Paris 7.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "forthcoming : Extraction out of NP and clitic-noun dependencies in French", "authors": [ { "first": "A", "middle": [], "last": "Abeille", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Abeille, forthcoming : Extraction out of NP and clitic-noun dependencies in French, in Abeille.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Tree\u2022adjoi11i11g Grammars. CSLI", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "A\" Rambow, 0. (eds.), Tree\u2022adjoi11i11g Grammars. CSLI, Stanford.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "forthcoming : Clitic Clirnbing and the power of TAG", "authors": [ { "first": "T", "middle": [], "last": "Bleam", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Bleam, forthcoming : Clitic Clirnbing and the power of TAG, in Abeille, A., Rambow. 0. (eds.).", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Can thc TAG derivation tree represent a semantic grnph", "authors": [ { "first": "M.-H", "middle": [], "last": "Candito", "suffix": "" }, { "first": "S", "middle": [], "last": "Kahane", "suffix": "" } ], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "M.-H. Candito, S. Kahane, 1998 ; Can thc TAG derivation tree represent a semantic grnph \"? An '!nswer in the light of the Meaning-Text Theory. This volume.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Syntactic Jocality and Trec Adjoining Grammar: Grammatical, Acquisition and Processing Perpectives", "authors": [ { "first": "R", "middle": [], "last": "Frank", "suffix": "" } ], "year": 1992, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. Frank, 1992 : Syntactic Jocality and Trec Adjoining Grammar: Grammatical, Acquisition and Processing Perpectives. Ph.D. thesis. Univ. of Pennsylvania.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Subjacency in a Trcc-Adjoining Grammar", "authors": [ { "first": "A", "middle": [], "last": "Kroch", "suffix": "" } ], "year": 1987, "venue": "A. Manaster-Ramcr. cd . Mathematics of la111", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Kroch, 1987 : Subjacency in a Trcc- Adjoining Grammar. In A. Manaster-Ramcr. cd . Mathematics of la111(ual(e.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Dependency Syntax : Theory and Practice", "authors": [ { "first": "", "middle": [], "last": "Mel'cuk", "suffix": "" } ], "year": 1988, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mel'cuk, 1988 : Dependency Syntax : Theory and Practice. Albany. State Univ . of New York Press.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "D-tree Grammars, ACL'95", "authors": [ { "first": "K", "middle": [], "last": "Rambow", "suffix": "" }, { "first": ".", "middle": [ "D" ], "last": "Vijay\u2022shanker", "suffix": "" }, { "first": "", "middle": [], "last": "Wcir", "suffix": "" } ], "year": 1995, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rambow, K. Vijay\u2022Shanker. D. Wcir. 1995 : D-tree Grammars, ACL'95.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "semanticeskom sinteze [On semantic svnthcsis J. Problemy kybernetiki, v", "authors": [ { "first": "A", "middle": [], "last": "Zolkovskij", "suffix": "" } ], "year": 1967, "venue": "Fr. Transl. In T.A. lnfonnations", "volume": "19", "issue": "0", "pages": "1--85", "other_ids": {}, "num": null, "urls": [], "raw_text": "A. Zolkovskij, 1. Mel'cuk, 1967 : 0 semanticeskom sinteze [On semantic svnthcsis J. Problemy kybernetiki, v. 19, 177-238. [Fr. Transl. In T.A. lnfonnations, 1970, #2, 1-85.]", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "uris": null, "num": null, "text": "Generalized adjunction Now consider the sentence:" }, "FIGREF1": { "type_str": "figure", "uris": null, "num": null, "text": "some two-rope elementary trees (for relative wh-words)" }, "FIGREF2": { "type_str": "figure", "uris": null, "num": null, "text": "Substitution of a two-rope d-tree" }, "FIGREF3": { "type_str": "figure", "uris": null, "num": null, "text": "shows the DG fcr (3b). To interprct a DG as ~ semantic graph, one needs to: \u2022 translate d-trees names into semantemcs \u2022 read substitution arcs as semantic dependcncies from the site of substitution to the suhstituted trcc: \u2022 read adjunction arcs as semantic dependcncics from the adjoined tree to the trce rccciving adjunction; \u2022 collapse some arcs that link coreferent node~. DTO). They are needed for instance to hlock extraction in the case of non-bridge verbs or to express constraints on double extractions and topicalization. MTT semantic graph l know the man whose car Peter borroll'ed" } } } }