Probabilistic Models for Learning a Semantic Parser Lexicon

Proceedings of NAACL-HLT 2016, pages 606–616, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Probabilistic Models for Learning a Semantic Parser Lexicon Jayant Krishnamurthy Allen Institute for Artificial Intelligence 2157 N. Northlake Way, Suite 110 Seattle, WA 98103 jayantk@allenai.org Abstract We introduce several probabilistic models for learning the lexicon of a semantic parser. Lex- icon learning is the first step of training a se- mantic parser for a new application domain and the quality of the learned lexicon signifi- cantly affects both the accuracy and efficiency of the final semantic parser. Existing work on lexicon learning has focused on heuris- tic methods that lack convergence guarantees and require significant human input in the form of lexicon templates or annotated logi- cal forms. In contrast, our probabilistic mod- els are trained directly from question/answer pairs using EM and our simplest model has a concave objective that guarantees convergence to a global optimum. An experimental evalu- ation on a set of 4th grade science questions demonstrates that our models improve seman- tic parser accuracy (35-70% error reduction) and efficiency (4-25x more sentences per sec- ond) relative to prior work despite using less human input. Our models also obtain compet- itive results on GEO880 without any dataset- specific engineering. 1 Introduction Semantic parsing has recently gained popularity as a technique for mapping from natural language to a formal meaning representation language, e.g., in or- der to answer questions against a database (Zelle and Mooney, 1993; Zettlemoyer and Collins, 2005). In order to train a semantic parser, one must first pro- vide a lexicon, which is a mapping from words in the language to statements in the meaning representa- tion language. This mapping defines the grammar of L λx.EATS(x, BASS) SKIP what ... λx.EATS(x, BASS) λf.λx.EATS(x, f) predator BASS SKIP of BASS bass Lexicon Entries: predator := N/NP : λf.λx.EATS(x, f) bass := NP : BASS Figure 1: Parse tree of a training example and the lexicon en- tries derived from it. the parser and thereby determines the set of meaning representations that can be produced for any given sentence. Therefore, a good lexicon is necessary to achieve both high accuracy and parsing speed. How- ever, the lexicon is unobserved in real semantic pars- ing applications, leading us to ask: how do we learn a lexicon for a semantic parser? This paper presents several novel probabilistic models for learning a semantic parser lexicon. Ex- isting lexicon learning algorithms are heuristic in na- ture and require either annotated logical forms or manually-specified lexicon entry templates during training. In contrast, our models do not require such templates and can be trained from question/answer pairs and other forms of weak supervision. Train- ing consists of optimizing an objective function with Expectation Maximization (EM), thereby guaran- 606 teeing convergence to a local optimum. Further- more, the objective function for our simplest model is concave, guaranteeing convergence to a global optimum. Our approach generates a probabilis- tic context-free grammar that represents the space of correct semantic …

Content truncated to 500 words