Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1223–1233, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification Tom Kwiatkowski∗ t.m.kwiatkowksi@sms.ed.ac.uk Luke Zettlemoyer† lsz@cs.washington.edu Sharon Goldwater∗ sgwater@inf.ed.ac.uk Mark Steedman∗ steedman@inf.ed.ac.uk ∗School of Informatics University of Edinburgh Edinburgh, EH8 9AB, UK †Computer Science & Engineering University of Washington Seattle, WA 98195 Abstract This paper addresses the problem of learn- ing to map sentences to logical form, given training data consisting of natural language sentences paired with logical representations of their meaning. Previous approaches have been designed for particular natural languages or specific meaning representations; here we present a more general method. The approach induces a probabilistic CCG grammar that represents the meaning of individual words and defines how these meanings can be com- bined to analyze complete sentences. We use higher-order unification to define a hy- pothesis space containing all grammars con- sistent with the training data, and develop an online learning algorithm that efficiently searches this space while simultaneously es- timating the parameters of a log-linear parsing model. Experiments demonstrate high accu- racy on benchmark data sets in four languages with two different meaning representations. 1 Introduction A key aim in natural language processing is to learn a mapping from natural language sentences to for- mal representations of their meaning. Recent work has addressed this problem by learning semantic parsers given sentences paired with logical meaning representations (Thompson & Mooney, 2002; Kate et al., 2005; Kate & Mooney, 2006; Wong & Mooney, 2006, 2007; Zettlemoyer & Collins, 2005, 2007; Lu et al., 2008). For example, the training data might consist of English sentences paired with lambda-calculus meaning representations: Sentence: which states border texas Meaning: λx.state(x) ∧ next to(x, tex) Given pairs like this, the goal is to learn to map new, unseen, sentences to their corresponding meaning. Previous approaches to this problem have been tailored to specific natural languages, specific mean- ing representations, or both. Here, we develop an approach that can learn to map any natural language to a wide variety of logical representations of lin- guistic meaning. In addition to data like the above, this approach can also learn from examples such as: Sentence: hangi eyaletin texas ye siniri vardir Meaning: answer(state(borders(tex))) where the sentence is in Turkish and the meaning representation is a variable-free logical expression of the type that has been used in recent work (Kate et al., 2005; Kate & Mooney, 2006; Wong & Mooney, 2006; Lu et al., 2008). The reason for generalizing to multiple languages is obvious. The need to learn over multiple repre- sentations arises from the fact that there is no stan- dard representation for logical form for natural lan- guage. Instead, existing representations are ad hoc, tailored to the application of interest. For example, the variable-free representation above was designed for building natural language interfaces to databases. Our approach works by inducing a combinatory categorial grammar (CCG) (Steedman, 1996, 2000). A CCG grammar …

Content truncated to 500 words