Accurate Unbounded Dependency Recovery using Generalized Categorial Grammars

Proceedings of COLING 2012: Technical Papers, pages 2125–2140, COLING 2012, Mumbai, December 2012. Accurate Unbounded Dependency Recovery using Generalized Categorial Grammars Luan NGUY EN 1 Marten VAN SCHIJN DEL2 William SCHU LER2 (1) University of Minnesota, Dept. of Computer Science and Engineering, Minneapolis, MN (2) The Ohio State University, Dept. of Linguistics, Columbus, OH lnguyen@cs.umn.edu, vanschm@ling.osu.edu, schuler@ling.osu.edu ABSTRACT Accurate recovery of predicate-argument dependencies is vital for interpretation tasks like information extraction and question answering, and unbounded dependencies may account for a significant portion of the dependencies in any given text. This paper describes a categorial grammar which, like other categorial grammars, imposes a small, uniform, and easily learnable set of semantic composition operations based on functor-argument relations, but like HPSG, is generalized to limit the number of categories used to those needed to enforce grammatical constraints. The paper also describes a novel reannotation system used to map existing resources based on Government and Binding Theory, like the Penn Treebank, into this categorial representation. This grammar is evaluated on an existing unbounded dependency recovery task (Rimell et al., 2009; Nivre et al., 2010). KEYWORDS: Grammar and formalisms, Semantics. 2125 1 Introduction Accurate recovery of predicate-argument dependencies is vital for interpretation tasks like information extraction and question answering, and unbounded dependencies may account for a significant portion of the dependencies in any given text. Many current interpretation models are based on PCFGs, trained on syntactic annotations from the Penn Treebank (Marcus et al., 1993). These often recover dependencies as a post-process to parsing, and often are not able to retrieve unbounded dependencies if they are optimized on syntactic representations that leave these dependencies out. Categorial grammars, on the other hand, have well-defined unbounded dependency representa- tions based on functor-argument relations in a small and easily-learnable set of composition operations. Such grammars — in particular, Combinatory Categorial Grammar (CCG) (Steed- man, 2000; Clark and Curran, 2007) — do well on unbounded dependency recovery tasks (Rimell et al., 2009) but not as well as models based on Head Driven Phrase-structure Grammar (HPSG) (Pollard and Sag, 1994; Miyao and Tsujii, 2005), given the same training. This may be attributed to implicit tradeoffs in many categorial frameworks that minimize the number of composition operations at the expense of large numbers of possible categories for each lexical item, which may lead to sparse data effects in training. HPSG models, in contrast, maintain a relatively large number of composition operations and a relatively small set of possible lexical categories, which are then used in a wider set of contexts. Can categorial grammars, which have well-studied semantic representations and are well suited for interpretation, obtain better performance on a general unbounded dependency extraction task if it adopts an HPSG-like strategy of re-using types in various contexts? This paper describes a categorial grammar which, like HPSG, is generalized to limit the number of categories used to those needed to enforce grammatical constraints, but like other categorial grammars, imposes a small, uniform, and easily learnable set of semantic composition operations based …

Content truncated to 500 words