Morpho-syntactic Lexical Generalization for CCG Semantic Parsing

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1284–1295, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Morpho-syntactic Lexical Generalization for CCG Semantic Parsing Adrienne Wang Computer Science & Engineering University of Washington Seattle, WA axwang@cs.washington.edu Tom Kwiatkowski Allen Institute for AI Seattle, WA tomk@allenai.org Luke Zettlemoyer Computer Science & Engineering University of Washington Seattle, WA lsz@cs.washington.edu Abstract In this paper, we demonstrate that significant performance gains can be achieved in CCG semantic parsing by introducing a linguistically moti- vated grammar induction scheme. We present a new morpho-syntactic fac- tored lexicon that models systematic variations in morphology, syntax, and semantics across word classes. The grammar uses domain-independent facts about the English language to restrict the number of incorrect parses that must be considered, thereby enabling effective learning from less data. Experiments in benchmark domains match previous models with one quarter of the data and provide new state-of-the-art results with all available data, including up to 45% relative test-error reduction. 1 Introduction Semantic parsers map sentences to formal representations of their meaning (Zelle and Mooney, 1996; Zettlemoyer and Collins, 2005; Liang et al., 2011). One common approach is to induce a probabilistic CCG grammar, which defines the meanings of individual words and phrases and how to best combine them to an- alyze complete sentences. There has been re- cent work developing learning algorithms for CCG semantic parsers (Kwiatkowski et al., 2010; Artzi and Zettlemoyer, 2011) and using them for applications ranging from question answering (Cai and Yates, 2013b; Kwiatkowski et al., 2013) to robot control (Matuszek et al., 2012; Krishnamurthy and Kollar, 2013). One key learning challenge for this style of learning is to induce the CCG lexicon, which lists possible meanings for each phrase and defines a set of possible parses for each sentence. Previous approaches have either hand-engineered a small set of lexi- cal templates (Zettlemoyer and Collins, 2005, 2007) or automatically learned such tem- plates (Kwiatkowski et al., 2010, 2011). These methods are designed to learn grammars that overgenerate; they produce spurious parses that can complicate parameter estimation. In this paper, we demonstrate that signif- icant gains can instead be achieved by using a more constrained, linguistically motivated grammar induction scheme. The grammar is restricted by introducing detailed syntac- tic modeling of a wider range of constructions than considered in previous work, for example introducing explicit treatments of relational nouns, Davidsonian events, and verb tense. We also introduce a new lexical generalization model that abstracts over systematic morpho- logical, syntactic, and semantic alternations within word classes. This includes, for exam- ple, the facts that verbs in relative clauses and nominal constructions (e.g., “flights departing NYC” and “departing flights”) should be in- finitival while verbs in phrases (e.g., “What flights depart at noon?”) should have tense. These grammar modeling techniques use uni- versal, domain-independent facts about the English language to restrict word usage to ap- propriate syntactic contexts, and as such are potentially applicable to any semantic parsing application. More specifically, …

Content truncated to 500 words