Building a Semantic Parser Overnight

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1332–1342, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Building a Semantic Parser Overnight Yushi Wang∗ Stanford University yushiw@cs.stanford.edu Jonathan Berant∗ Stanford University joberant@stanford.edu Percy Liang Stanford University pliang@cs.stanford.edu Abstract How do we build a semantic parser in a new domain starting with zero training ex- amples? We introduce a new methodol- ogy for this setting: First, we use a simple grammar to generate logical forms paired with canonical utterances. The logical forms are meant to cover the desired set of compositional operators, and the canon- ical utterances are meant to capture the meaning of the logical forms (although clumsily). We then use crowdsourcing to paraphrase these canonical utterances into natural utterances. The resulting data is used to train the semantic parser. We fur- ther study the role of compositionality in the resulting paraphrases. Finally, we test our methodology on seven domains and show that we can build an adequate se- mantic parser in just a few hours. 1 Introduction By mapping natural language utterances to exe- cutable logical forms, semantic parsers have been useful for a variety of applications requiring pre- cise language understanding (Zelle and Mooney, 1996; Zettlemoyer and Collins, 2005; Liang et al., 2011; Berant et al., 2013; Kwiatkowski et al., 2013; Artzi and Zettlemoyer, 2013; Kushman and Barzilay, 2013). Previous work has focused on how to train a semantic parser given input utter- ances, but suppose we wanted to build a seman- tic parser for a new domain—for example, a natu- ral language interface into a publications database. Since no such interface exists, we do not even have a naturally occurring source of input utterances that we can annotate. So where do we start? In this paper, we advocate a functionality- driven process for rapidly building a semantic ∗ Both authors equally contributed to the paper. Domain Seed lexicon article → TypeNP[article] publication date → RelNP[publicationDate] cites → VP/NP[cites] ... Logical forms and canonical utterances article that has the largest publication date argmax(type.article, publicationDate) person that is author of the most number of article argmax(type.person, R(λx.count(type.article ⊓ author.x))) ... Paraphrases what is the newest published article? who has published the most articles? ... Semantic parser (1) by builder (∼30 minutes) (2) via domain-general grammar (3) via crowdsourcing (∼5 hours) (4) by training a paraphrasing model Figure 1: Functionality-driven process for build- ing semantic parsers. The two red boxes are the domain-specific parts provided by the builder of the semantic parser, and the other two are gener- ated by the framework. parser in a new domain. At a high-level, we seek to minimize the amount of work needed for a new domain by factoring out the domain- general aspects (done by our framework) from the domain-specific ones (done by the builder of the semantic parser). We assume that the builder already has the desired functionality of the semantic parser in mind—e.g., the …

Content truncated to 500 words