An Alternative to Head-Driven Approaches for Parsing a (Relatively) Free Word-Order Language

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 842–851, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP An Alternative to Head-Driven Approaches for Parsing a (Relatively) Free Word-Order Language Reut Tsarfaty Khalil Sima’an Remko Scha Institute for Logic Language and Computation University of Amsterdam {r.tsarfaty,k.simaan,r.scha}@uva.nl Abstract Applying statistical parsers developed for English to languages with freer word- order has turned out to be harder than expected. This paper investigates the adequacy of different statistical parsing models for dealing with a (relatively) free word-order language. We show that the recently proposed Relational- Realizational (RR) model consistently outperforms state-of-the-art Head-Driven (HD) models on the Hebrew Treebank. Our analysis reveals a weakness of HD models: their intrinsic focus on configu- rational information. We conclude that the form-function separation ingrained in RR models makes them better suited for pars- ing nonconfigurational phenomena. 1 Introduction Parsing technology has come a long way since Charniak (1996) demonstrated that a simple tree- bank PCFG performs better than any other parser (with F175 accuracy) on parsing the WSJ Penn treebank (Marcus et al., 1993). Treebank Gram- mars (Scha, 1990; Charniak, 1996) trained on large corpora nowadays present the best available means to parse natural language text. The performance curve for parsing the WSJ was a steep one at first, as the incorporation of no- tions such as head, distance, subcategorization (Charniak, 1997; Collins, 1999) brought about a dramatic increase in parsing accuracy to the level of F188. Discriminative approaches, Data- Oriented Parsing (‘all-subtrees’) approaches, and self-training techniques brought further improve- ments, and recent results are starting to level off at around F192.1 (McClosky et al., 2008). As the interest of the NLP community grows to encompass more languages, we observe efforts towards adapting an English parser for parsing other languages (e.g., (Collins et al., 1999)), or towards designing a language-independent frame- work based on principles underlying the mod- els for parsing English (Bikel, 2002). The per- formance curve for parsing other languages with these models looks rather different. A case in point is Modern Standard Arabic. Since the initial ef- fort of (Bikel, 2002) to parse the Arabic treebank (Maamouri et al., 2004), which yielded F175 ac- curacy, four years and successive revisions have led to no more than F179 (Maamouri et al., 2008). This pattern from Arabic is not peculiar. The level of state-of-the-art results for other languages still lags behind those for English, even after putting considerable effort into the adaptation.1 Given that these languages are inherently differ- ent from English and from one another, it appears that we cannot avoid a question concerning the ad- equacy of the models used to parse them. That is, given the properties of a language, which model- ing strategy would be appropriate for parsing it? Until recently, there has been practically no computationally affordable alternative to the Head-Driven (HD) approach in the development of phrase-structure based statistical parsing mod- els. Recently, we proposed the Relational- Realizational (RR) approach that rests upon differ- ent premises …

Content truncated to 500 words