Long-Distance Dependency Resolution in Automatically Acquired Wide-Coverage PCFG-Based LFG Approximations

Long-Distance Dependency Resolution in Automatically Acquired Wide-Coverage PCFG-Based LFG Approximations Aoife Cahill, Michael Burke, Ruth O’Donovan, Josef van Genabith, Andy Way National Centre for Language Technology and School of Computing, Dublin City University, Dublin, Ireland {acahill,mburke,rodonovan,josef,away}@computing.dcu.ie Abstract This paper shows how finite approximations of long distance dependency (LDD) resolution can be obtained automatically for wide-coverage, robust, probabilistic Lexical-Functional Grammar (LFG) resources acquired from treebanks. We extract LFG subcategorisation frames and paths linking LDD reentrancies from f-structures generated automati- cally for the Penn-II treebank trees and use them in an LDD resolution algorithm to parse new text. Unlike (Collins, 1999; Johnson, 2002), in our ap- proach resolution of LDDs is done at f-structure (attribute-value structure representations of basic predicate-argument or dependency structure) with- out empty productions, traces and coindexation in CFG parse trees. Currently our best automatically induced grammars achieve 80.97% f-score for f- structures parsing section 23 of the WSJ part of the Penn-II treebank and evaluating against the DCU 1051 and 80.24% against the PARC 700 Depen- dency Bank (King et al., 2003), performing at the same or a slightly better level than state-of-the-art hand-crafted grammars (Kaplan et al., 2004). 1 Introduction The determination of syntactic structure is an im- portant step in natural language processing as syn- tactic structure strongly determines semantic inter- pretation in the form of predicate-argument struc- ture, dependency relations or logical form. For a substantial number of linguistic phenomena such as topicalisation, wh-movement in relative clauses and interrogative sentences, however, there is an im- portant difference between the location of the (sur- face) realisation of linguistic material and the loca- tion where this material should be interpreted se- mantically. Resolution of such long-distance de- pendencies (LDDs) is therefore crucial in the de- termination of accurate predicate-argument struc- 1Manually constructed f-structures for 105 randomly se- lected trees from Section 23 of the WSJ section of the Penn-II Treebank ture, deep dependency relations and the construc- tion of proper meaning representations such as log- ical forms (Johnson, 2002). Modern unification/constraint-based grammars such as LFG or HPSG capture deep linguistic infor- mation including LDDs, predicate-argument struc- ture, or logical form. Manually scaling rich uni- fication grammars to naturally occurring free text, however, is extremely time-consuming, expensive and requires considerable linguistic and computa- tional expertise. Few hand-crafted, deep unification grammars have in fact achieved the coverage and robustness required to parse a corpus of say the size and complexity of the Penn treebank: (Riezler et al., 2002) show how a deep, carefully hand-crafted LFG is successfully scaled to parse the Penn-II tree- bank (Marcus et al., 1994) with discriminative (log- linear) parameter estimation techniques. The last 20 years have seen continuously increas- ing efforts in the construction of parse-annotated corpora. Substantial treebanks2 are now available for many languages (including English, Japanese, Chinese, German, French, Czech, Turkish), others are currently under construction (Arabic, Bulgarian) or near completion (Spanish, Catalan). Treebanks have been enormously influential in the develop- ment of robust, state-of-the-art parsing technology: grammars (or grammatical information) automat- ically …

Content truncated to 500 words