Grounded Unsupervised Semantic Parsing

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 933–943, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Grounded Unsupervised Semantic Parsing Hoifung Poon One Microsoft Way Microsoft Research Redmond, WA 98052, USA hoifung@microsoft.com Abstract We present the first unsupervised ap- proach for semantic parsing that rivals the accuracy of supervised approaches in translating natural-language questions to database queries. Our GUSP system produces a semantic parse by annotat- ing the dependency-tree nodes and edges with latent states, and learns a proba- bilistic grammar using EM. To compen- sate for the lack of example annotations or question-answer pairs, GUSP adopts a novel grounded-learning approach to leverage database for indirect supervision. On the challenging ATIS dataset, GUSP attained an accuracy of 84%, effectively tying with the best published results by su- pervised approaches. 1 Introduction Semantic parsing maps text to a formal mean- ing representation such as logical forms or struc- tured queries. Recently, there has been a bur- geoning interest in developing machine-learning approaches for semantic parsing (Zettlemoyer and Collins, 2005; Zettlemoyer and Collins, 2007; Mooney, 2007; Kwiatkowski et al., 2011), but the predominant paradigm uses supervised learn- ing, which requires example annotations that are costly to obtain. More recently, several grounded- learning approaches have been proposed to alle- viate the annotation burden (Chen and Mooney, 2008; Kim and Mooney, 2010; B¨orschinger et al., 2011; Clarke et al., 2010; Liang et al., 2011). In particular, Clarke et al. (2010) and Liang et al. (2011) proposed methods to learn from question- answer pairs alone, which represents a significant advance. However, although these methods exon- erate annotators from mastering specialized logi- cal forms, finding the answers for complex ques- tions still requires non-trivial effort. 1 Poon & Domingos (2009, 2010) proposed the USP system for unsupervised semantic parsing, which learns a parser by recursively clustering and composing synonymous expressions. While their approach completely obviates the need for di- rect supervision, their target logic forms are self- induced clusters, which do not align with existing database or ontology. As a result, USP can not be used directly to answer complex questions against an existing database. More importantly, it misses the opportunity to leverage database for indirect supervision. In this paper, we present the GUSP system, which combines unsupervised semantic parsing with grounded learning from a database. GUSP starts with the dependency tree of a sentence and produces a semantic parse by annotating the nodes and edges with latent semantic states derived from the database. Given a set of natural-language questions and a database, GUSP learns a prob- abilistic semantic grammar using EM. To com- pensate for the lack of direct supervision, GUSP constrains the search space using the database schema, and bootstraps learning using lexical scores computed from the names and values of database elements. Unlike previous grounded-learning approaches, GUSP does not require ambiguous annotations or oracle answers, but rather focuses on lever- aging database contents that are readily avail- able. Unlike USP, GUSP predetermines the tar- get …

Content truncated to 500 words