{ "paper_id": "W98-0137", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T06:03:12.524797Z" }, "title": "Exploiting Semantic Dependencies in Parsing", "authors": [ { "first": "William", "middle": [], "last": "Schuler", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Pennsylvania Philadelphia", "location": { "postCode": "19103", "region": "PA" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In this paper we describe a semantic dependency model for estimating probabilities in a stochastic TAG parser (Resnik, 1992) (Schabes, 1992), and we compare it with the syntactic dependency model inherent in a TAG derivation using the flat treatment of modifiers described in (Schabes and Shieber, 1994).", "pdf_parse": { "paper_id": "W98-0137", "_pdf_hash": "", "abstract": [ { "text": "In this paper we describe a semantic dependency model for estimating probabilities in a stochastic TAG parser (Resnik, 1992) (Schabes, 1992), and we compare it with the syntactic dependency model inherent in a TAG derivation using the flat treatment of modifiers described in (Schabes and Shieber, 1994).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The use of syntactic dependencies to estimate parser probabilities is not uncommon (Eisner, 1996) (Collins, 1997) (Charniak, 1997) . Typically, a maximum probability parse is estimated from bigram statistics of lexical items that participate in head-modifier or head-complement dependencies with other lexical items. These dependencies can be characterized as ( head, label, modifier ) triples and ( head, label, complement ) triples -or as labeled directed arcs in a graph -which have the property that each lexical item may participate as a modifier or a complement in no more than one dependency. Using a TAG derivation tree (Joshi, 1987) with a flat treatment of modifiers (Schabes and Shieber, 1994) , it is possible to capture the long distance dependencies of wh-extractions and relative clauses as adjacent arcs in a dependency structure, making them available for probability estimates withiu the parser as well. In this case, the head-complement dependencies for a sentence correspond to a set S of substitution triples (/, rJ, a) (where tree a substitutes into tree 'Y at note address ?J), and the head-modifier dependencies correspond to a set A of adjunction triples (/, 1}; \u00df) (where tree \u00df adjoins into tree 'Y at node address 7J), in a probabilistic TAG (Resnik, 1992) .1", "cite_spans": [ { "start": 83, "end": 97, "text": "(Eisner, 1996)", "ref_id": "BIBREF3" }, { "start": 98, "end": 113, "text": "(Collins, 1997)", "ref_id": "BIBREF2" }, { "start": 114, "end": 130, "text": "(Charniak, 1997)", "ref_id": "BIBREF1" }, { "start": 628, "end": 641, "text": "(Joshi, 1987)", "ref_id": "BIBREF4" }, { "start": 677, "end": 704, "text": "(Schabes and Shieber, 1994)", "ref_id": "BIBREF8" }, { "start": 1270, "end": 1284, "text": "(Resnik, 1992)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Although the TAG-based syntactic dependency rnodel has the necessary domain of locality (in terms of adjacent arcs on the derivation tree) to accurately guide a statistical parser, it is still susceptible to sparse data effects, in part because it does not generalize attachment statistics across syntactic transformations. An adjective used as a declarative predicate, for example, could not draw on attachment statistics for the same adjective used as a modifier, or as a predicate in a relative clause, and vice versa, because each transformation uses a different syntactic dependency structure. The triples in the syntactic dependency sets S and A for the sentences, 11 The damaged handle is attached to the drawer,\" and {CThe handle attached to the drawer is damaged,\" are represented as arcs in Figure 1 .", "cite_spans": [ { "start": 671, "end": 673, "text": "11", "ref_id": null } ], "ref_spans": [ { "start": 801, "end": 809, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In order to group these attachment statistics into denser pools of data, we need to abstract a common semantic structure from the various syntactic structures, effectively adopting a common argument frame for each transformation. This means that each auxiliary tree must have an argument position corresponding to the subject substitution site in its predicative transformation if it is a modifier auxiliary, or corresponding to the wh-object substitution site in its object-extraction transformation if it is a predicative auxiliary. 2 For convention, we place 1 Although Resnik uses a direct function S(-y, 11, a) to the [O -1) interval where we use a probability of set membership 'P((\"{, fJ, a) E S). Also note that this correspondence between head-complement dependencies and substitution dependencies is not strictly true in the case of predicative auxiliaries (Schabes and Shieber, 1994) , which are handled by adjunction in TAG.", "cite_spans": [ { "start": 867, "end": 894, "text": "(Schabes and Shieber, 1994)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "attach to ~ handle door NPtod 'V darnage", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The damaged handle is attached to the door.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The handle attached to the door is damaged. Figure 1 : Syntactic dependencies in TAG this extra argument position at the foot node of the auxiliary tree, so the auxiliary takes the tree it adjoins into as an argument. This means that our semantic dependency model effectively reverses the direction of dependencies involved in adjunction from the syntactic model.The triples in the semantic dependency set 1) for the sentences, 11 The damaged handle is attached to the drawer,\" and \"The handle attached to the drawer is damaged,\" are represented as arcs in Figure 2 . Formally, we augment the syntactic dependency sets S and A with a semantic dependency set V of ( predicate, label, argument ) triples defined as follows:", "cite_spans": [ { "start": 428, "end": 430, "text": "11", "ref_id": null } ], "ref_spans": [ { "start": 44, "end": 52, "text": "Figure 1", "ref_id": null }, { "start": 557, "end": 565, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For every substitution (head-complement) dependency ('Y, 71, a) in S add a predicateargument dependency (anchar(ry), argnum (J, 17) , anchor(a)} to V; and \u2022 For every adjunction (head-modifier) dependency", "cite_spans": [ { "start": 126, "end": 133, "text": "(J, 17)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "('Y, 17, \u00df) in A add a predicate-argument dependency (anchar(\u00df), argnum(\u00df, f oot(\u00df)), anchar('Y)) to V;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "where anchor(a) returns the lexical and10r of tree a, and argnum(a,17} returns the semantic argument position corresponding to node 17 in tree a. In this way we can combine argument attachrnent distributions for initial tree trans-\u2022 formations and auxiliary tree transformations into a common attachment distribution for the underlying predicate. Parsing proceeds in three passes of O(n 6 ) complexity. First, the chart is filled in from the bottom up, as described in (Schabes et al., 1988) , and the input is recognized or rejected. The parser then constructs a shared forest (Vijay-Shanker and Weir, 1993) top-down from the elements in the chart, ignoring those items on bottom-up dead ends. Finally, the parser proceeds with the more expensive Operations of feature unification and probability estimation on the reduced set of nodes in the shared forest. The chart consists of a set of items that each specify a node address 77 in an elementary tree a, a top (T) or bottom (.l) marker denoting the phase of operation on the node, and four indices i,j,k, and l, composing the extent of the node's coverage in the sentence: (o:, 77, T 1 i, j, k, l). The shared forest consists of an and/or graph, with 'or' arcs from each non-dead-end chart item to instantiations of the parser productions that could have produced it, and 'and' arcs from each instantiation of a parser production to the chart items it would have required.", "cite_spans": [ { "start": 469, "end": 491, "text": "(Schabes et al., 1988)", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In order to select a most-preferred parse for an ambiguous input, a highest-probability item is selected from the top node in the shared forest, and a parse is read off below it by traversing the subordinate items with the most probable dependencies. The probability of each shared forest item is computed as the maximurn of the probabilities of its 'or'-adjacent parser productions. The probability of each instantiation of a parser production is cornputed as the proba-bility of the relevant dependency for that production multiplied by the probabilities of the chart items that production required. Finally, the probability of each parse must be multiplied by the probability of each elementary tree given a lexical item in the input.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The probability model is adapted from (Resnik, 1992) , which assigns a probability to any arc (a, 7] , \u00df} (where tree \u00df is attached to tree a at node address 77) being in the set of substitutions S or adjunctions A in a derivation.", "cite_spans": [ { "start": 38, "end": 52, "text": "(Resnik, 1992)", "ref_id": "BIBREF5" }, { "start": 94, "end": 100, "text": "(a, 7]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The root of the derivation tree is represented as (MAIN, O, a} in S, and null adjunctions (which terminate the adjunction of modifiers at a node)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "are represented as (a, 7] , t} in A. Finally, the probability of a tree a is represented as the probability of the double {anchor(a), tree(a)) being in the set r of elementary trees used in a parse.", "cite_spans": [ { "start": 19, "end": 25, "text": "(a, 7]", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Probabilities for the dependencies in a parser production are estimated from observed frequencies that a child predicate c (the base-form anchor of a tree) occurs in argument position a of a parent predicate p (the base-form anchor of another tree), within some training set", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "'D of dependency structures: F( (a, p, c) E 'D).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The top-level dependency is represented in 'D as (MAIN, O, c) , and null adjunctions are represented as (NULL, 0, c) . 3 Note that we use the same dependencies as Resnik (the syntactic dependency sets Sand A) in describing the probability model, and use the semantic dependencies ('D) only in the estimation of those probabilities.", "cite_spans": [ { "start": 104, "end": 110, "text": "(NULL,", "ref_id": null }, { "start": 111, "end": 113, "text": "0,", "ref_id": null }, { "start": 114, "end": 116, "text": "c)", "ref_id": null } ], "ref_spans": [ { "start": 49, "end": 61, "text": "(MAIN, O, c)", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Probabilities are estimated as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For any topmost item in a derivation tree: (a,O, T,0,-,-,n} the initial probability would be:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "'P({MAIN,O,a) ES 1 (MAIN,O, _)ES)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "which we estimate as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F((M AI N,O,anchor(o:))E'D) F( (MAI N,0,-)ED)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For any chart production for the substitution of initiai tree a into / at node address 17, where i and j are indices, and 17 is a substitution site in 'Y with the same label as the root of a:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "(a,O, T,i,-,-,j) ('y, 17, T, i, -, -, j) the probability would be:", "cite_spans": [ { "start": 17, "end": 40, "text": "('y, 17, T, i, -, -, j)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "'P( ('y, 7] 1 a} E S 1 ('y, 7], -} E S)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "which we estimate as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F (( a'!chor( 'Y) ,a rgnum( 'Y ,ry) ,anchor ( o: )}E'D)", "cite_spans": [ { "start": 2, "end": 17, "text": "(( a'!chor( 'Y)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F ( (anchor('Y) ,argnum( -y,77) 1 -}E'D)", "cite_spans": [ { "start": 2, "end": 15, "text": "( (anchor('Y)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For any chart production for adjunction of auxiliary tree \u00df into 'Y at node address 77, where i,j,i 1 ,j 1 ,p and q are indices, and 17 is an adjunction site in/' with the same label as the root of \u00df; ('Y,7J,1-,i1,p,q,j1} (\u00df,O, T,i,i',j',j) (!', 77, 1-, i, p, q, j) the probability would be:", "cite_spans": [ { "start": 203, "end": 242, "text": "('Y,7J,1-,i1,p,q,j1} (\u00df,O, T,i,i',j',j)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "'P((\"f, 17,\u00df) E A 1 {'Y, 7J, -) E A)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "which we estimate as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F({anchor(\u00df),~rgnum(\u00df,foot(\u00df)),anchor(-y))E1>) F( (_,_,anchor( -y))E'D)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For any chart production for closing adjunction at a node address 17 in tree 'Y:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "('Y,7], T,i,j,k,l) ()\" 17, 1-, i, j, k, l)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "the probability would be:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "'P(('Y,7J,\u20ac) E A j ('Y,77,_) E A)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "which we estimate as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F{(NU LL,O,anchor('Y))E'D)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Fe (-,-,anchor('Y ) )ED)", "cite_spans": [ { "start": 3, "end": 19, "text": "(-,-,anchor('Y )", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 For any other chart production, the probability would be l.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Finally, the probability that each elementary tree a is in the set of trees r used in the parse, given a lexical item is:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "P((anchor(a), tree(a)) E Tl(anchor(a), _)ET)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "which we estimate as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "F((anchor( o) ,tree( o))ET) F((anchor(o:},-)ET) 3 Practical Issues", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The extended goal of this project was to provide a natural language interface for \"Jack\" (Badler et al., 1993) , a human-like agent that answers questions and carries out instructions in a virtual 3-D environment. The system's restricted domain makes unknown words and unknown syntactic structures unlikely, and th~ goal of translating inputs into a formal language for the agent avoids the