Relational-Realizational Parsing

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 889–896 Manchester, August 2008 Relational-Realizational Parsing Reut Tsarfaty and Khalil Sima’an Institute for Logic, Language and Computation, University of Amsterdam Plantage Muidergracht 24, 1018TV, Amsterdam, The Netherlands {rtsarfat,simaan}@science.uva.nl Abstract State-of-the-art statistical parsing models applied to free word-order languages tend to underperform compared to, e.g., pars- ing English. Constituency-based mod- els often fail to capture generalizations that cannot be stated in structural terms, and dependency-based models employ a ‘single-head’ assumption that often breaks in the face of multiple exponence. In this paper we suggest that the position of a con- stituent is a form manifestation of its gram- matical function, one among various pos- sible means of realization. We develop the Relational-Realizational approach to pars- ing in which we untangle the projection of grammatical functions and their means of realization to allow for phrase-structure variability and morphological-syntactic in- teraction. We empirically demonstrate the application of our approach to pars- ing Modern Hebrew, obtaining 7% error reduction from previously reported results. 1 Introduction Many broad-coverage statistical parsers to date are constituency-based with a Probabilistic Context- Free Grammar (PCFG) or a Stochastic Tree Sub- stitution Grammar (STSG) at their backbone. The majority of such models belong to a Head-Driven paradigm, in which a head constituent is gen- erated first, providing a positional anchor for subsequent (e.g., Markovian) sisters’ generation. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Constituency-based models, lexicalized and un- lexicalized alike, demonstrate state-of-the-art per- formance for parsing English (Charniak, 1997; Collins, 2003; Klein and Manning, 2003; Bod, 2003), yet a direct application of such models to parsing less configurational languages often fails to yield comparable results. The parameters of such parsers capture generalizations that are eas- ily stated in structural terms (e.g., subjects linearly precede predicates, VPs dominate objects, etc.) which may not be adequate for parsing languages with less configurational character. A different vein of research explores data-driven dependency-based parsing methods (e.g., (Mc- Donald et al., 2005)) which seem to be intuitively more adequate for the task. It turns out, how- ever, that even such models fail to provide the desired remedy. Recent reports by (Nivre, 2007) delineated a class of richly-inflected languages with relatively free word-order (including Greek, Basque, and Modern Standard Arabic) for which the parsers performed poorly, regardless of the parsing method used. The need for parsing meth- ods that can effectively cope with such phenomena doesn’t seem to have been eliminated by depen- dency parsing — perhaps quite the contrary. The essential argument we promote here is that in order to deal with the kind of variation that is empirically observed cross-linguistically an al- ternative view of the generation process is re- quired. Our Relational-Realizational parsing pro- posal, strongly inspired by Relational Grammar (Perlmutter, 1982), takes grammatical relations such as ‘Subject’ and ‘Predicate’ as central, primi- tive notions of the syntactic representation, and re- tains a distinction between the projection of such relations and …

Content truncated to 500 words