Learning Compact Lexicons for CCG Semantic Parsing

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1273–1283, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Learning Compact Lexicons for CCG Semantic Parsing Yoav Artzi∗ Computer Science & Engineering University of Washington Seattle, WA 98195 yoav@cs.washington.edu Dipanjan Das Slav Petrov Google Inc. 76 9th Avenue New York, NY 10011 {dipanjand,slav}@google.com Abstract We present methods to control the lexicon size when learning a Combinatory Cate- gorial Grammar semantic parser. Existing methods incrementally expand the lexicon by greedily adding entries, considering a single training datapoint at a time. We pro- pose using corpus-level statistics for lexi- con learning decisions. We introduce vot- ing to globally consider adding entries to the lexicon, and pruning to remove entries no longer required to explain the training data. Our methods result in state-of-the-art performance on the task of executing se- quences of natural language instructions, achieving up to 25% error reduction, with lexicons that are up to 70% smaller and are qualitatively less noisy. 1 Introduction Combinatory Categorial Grammar (Steedman, 1996, 2000, CCG, henceforth) is a commonly used formalism for semantic parsing – the task of mapping natural language sentences to for- mal meaning representations (Zelle and Mooney, 1996). Recently, CCG semantic parsers have been used for numerous language understanding tasks, including querying databases (Zettlemoyer and Collins, 2005), referring to physical objects (Ma- tuszek et al., 2012), information extraction (Kr- ishnamurthy and Mitchell, 2012), executing in- structions (Artzi and Zettlemoyer, 2013b), gen- erating regular expressions (Kushman and Barzi- lay, 2013), question-answering (Cai and Yates, 2013) and textual entailment (Lewis and Steed- man, 2013). In CCG, a lexicon is used to map words to formal representations of their meaning, which are then combined using bottom-up opera- tions. In this paper we present learning techniques ∗This research was carried out at Google. chair ⊢ N : λx.chair(x) chair ⊢ N : λx.sofa(x) chair ⊢ AP : λa.len(a, 3) chair ⊢ NP : A(λx.corner(x)) chair ⊢ ADJ : λx.hall(x) Figure 1: Lexical entries for the word chair as learned with no corpus-level statistics. Our approach is able to correctly learn only the top two bolded entries. to explicitly control the size of the CCG lexicon, and show that this results in improved task perfor- mance and more compact models. In most approaches for inducing CCGs for se- mantic parsing, lexicon learning and parameter es- timation are performed jointly in an online algo- rithm, as introduced by Zettlemoyer and Collins (2007). To induce the lexicon, words extracted from the training data are paired with CCG cat- egories one sample at a time (for an overview of CCG, see §2). Joint approaches have the potential advantage that only entries participating in suc- cessful parses are added to the lexicon. However, new entries are added greedily and these decisions are never revisited at later stages. In practice, this often results in a large and noisy lexicon. Figure 1 lists a sample of CCG lexical entries learned for the word chair with a greedy …

Content truncated to 500 words