Recovering Non-Local Dependencies for Chinese

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 257–266, Prague, June 2007. c⃝2007 Association for Computational Linguistics Recovering Non-Local Dependencies for Chinese Yuqing Guo NCLT, School of Computing Dublin City University Dublin 9, Ireland yguo@computing.dcu.ie Haifeng Wang Toshiba (China) Research and Development Center Beijing, 100738, China wanghaifeng@rdc.toshiba.com.cn Josef van Genabith NCLT, School of Computing Dublin City University IBM CAS, Dublin, Ireland josef@computing.dcu.ie Abstract To date, work on Non-Local Dependencies (NLDs) has focused almost exclusively on English and it is an open research question how well these approaches migrate to other languages. This paper surveys non-local de- pendency constructions in Chinese as repre- sented in the Penn Chinese Treebank (CTB) and provides an approach for generating proper predicate-argument-modifier struc- tures including NLDs from surface context- free phrase structure trees. Our approach re- covers non-local dependencies at the level of Lexical-Functional Grammar f-structures, using automatically acquired subcategorisa- tion frames and f-structure paths linking an- tecedents and traces in NLDs. Currently our algorithm achieves 92.2% f-score for trace insertion and 84.3% for antecedent recovery evaluating on gold-standard CTB trees, and 64.7% and 54.7%, respectively, on CTB- trained state-of-the-art parser output trees. 1 Introduction A substantial number of linguistic phenomena such as topicalisation, relativisation, coordination and raising & control constructions, permit a constituent in one position to bear the grammatical role asso- ciated with another position. These relationships are referred to Non-Local Dependencies (NLDs), where the surface location of the constituent is called /antecedent0, and the site where the an- tecedent should be interpreted semantically is called /trace0. Capturing non-local dependencies is cru- cial to the accurate and complete determination of semantic interpretation in the form of predicate- argument-modifier structures or deep dependencies. However, with few exceptions (Model 3 of Collins, 1999; Schmid, 2006), output trees pro- duced by state-of-the-art broad coverage statistical parsers (Charniak, 2000; Bikel, 2004) are only sur- face context-free phrase structure trees (CFG-trees) without empty categories and coindexation to repre- sent displaced constituents. Because of the impor- tance of non-local dependencies in the proper de- termination of predicate-argument structures, recent years have witnessed a considerable amount of re- search on reconstructing such hidden relationships in CFG-trees. Three strategies have been proposed: (i) post-processing parser output with pattern match- ers (Johnson, 2002), linguistic principles (Campbell, 2004) or machine learning methods (Higgins, 2003; Levy and Manning, 2004; Gabbard et al., 2006) to recover empty nodes and identify their antecedents;1 (ii) integrating non-local dependency recovery into the parser by enriching a simple PCFG model with GPSG-style gap features (Collins, 1999; Schmid, 2006); (iii) pre-processing the input sentence with a finite-state trace tagger which detects empty nodes before parsing, and identify the antecedents on the parser output with the gap information (Dienes and Dubey, 2003a; Dienes and Dubey, 2003b). In addition to CFG-oriented approaches, a num- ber of richer treebank-based grammar acquisition and parsing methods based on HPSG (Miyao et al., 2003), CCG (Clark and Hockenmaier, 2002), LFG (Riezler et al., 2002; Cahill …

Content truncated to 500 words