Compositional Semantic Parsing on Semi-Structured Tables

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1470–1480, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Compositional Semantic Parsing on Semi-Structured Tables Panupong Pasupat Computer Science Department Stanford University ppasupat@cs.stanford.edu Percy Liang Computer Science Department Stanford University pliang@cs.stanford.edu Abstract Two important aspects of semantic pars- ing for question answering are the breadth of the knowledge source and the depth of logical compositionality. While existing work trades off one aspect for another, this paper simultaneously makes progress on both fronts through a new task: answering complex questions on semi-structured ta- bles using question-answer pairs as super- vision. The central challenge arises from two compounding factors: the broader do- main results in an open-ended set of re- lations, and the deeper compositionality results in a combinatorial explosion in the space of logical forms. We propose a logical-form driven parsing algorithm guided by strong typing constraints and show that it obtains significant improve- ments over natural baselines. For evalua- tion, we created a new dataset of 22,033 complex questions on Wikipedia tables, which is made publicly available. 1 Introduction In semantic parsing for question answering, nat- ural language questions are converted into logi- cal forms, which can be executed on a knowl- edge source to obtain answer denotations. Early semantic parsing systems were trained to answer highly compositional questions, but the knowl- edge sources were limited to small closed-domain databases (Zelle and Mooney, 1996; Wong and Mooney, 2007; Zettlemoyer and Collins, 2007; Kwiatkowski et al., 2011). More recent work sacrifices compositionality in favor of using more open-ended knowledge bases such as Freebase (Cai and Yates, 2013; Berant et al., 2013; Fader et al., 2014; Reddy et al., 2014). However, even these broader knowledge sources still define a Year City Country Nations 1896 Athens Greece 14 1900 Paris France 24 1904 St. Louis USA 12 . . . . . . . . . . . . 2004 Athens Greece 201 2008 Beijing China 204 2012 London UK 204 x1: “Greece held its last Summer Olympics in which year?” y1: {2004} x2: “In which city’s the first time with at least 20 nations?” y2: {Paris} x3: “Which years have the most participating countries?” y3: {2008, 2012} x4: “How many events were in Athens, Greece?” y4: {2} x5: “How many more participants were there in 1900 than in the first year?” y5: {10} Figure 1: Our task is to answer a highly composi- tional question from an HTML table. We learn a semantic parser from question-table-answer triples {(xi, ti, yi)}. rigid schema over entities and relation types, thus restricting the scope of answerable questions. To simultaneously increase both the breadth of the knowledge source and the depth of logical compositionality, we propose a new task (with an associated dataset): answering a question using an HTML table as the knowledge source. Figure 1 shows several question-answer pairs and an ac- companying table, which are typical of those in …

Content truncated to 500 words