Neural Semantic Parsing with Type Constraints for Semi-Structured Tables

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1516–1526 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Neural Semantic Parsing with Type Constraints for Semi-Structured Tables Jayant Krishnamurthy,1 Pradeep Dasigi,2 and Matt Gardner1 1Allen Institute for Artificial Intelligence 2Carnegie Mellon University {jayantk, mattg}@allenai.org, pdasigi@cs.cmu.edu Abstract We present a new semantic parsing model for answering compositional questions on semi-structured Wikipedia tables. Our parser is an encoder-decoder neural net- work with two key technical innovations: (1) a grammar for the decoder that only generates well-typed logical forms; and (2) an entity embedding and linking mod- ule that identifies entity mentions while generalizing across tables. We also in- troduce a novel method for training our neural model with question-answer super- vision. On the WIKITABLEQUESTIONS data set, our parser achieves a state-of-the- art accuracy of 43.3% for a single model and 45.9% for a 5-model ensemble, im- proving on the best prior score of 38.7% set by a 15-model ensemble. These re- sults suggest that type constraints and en- tity linking are valuable components to in- corporate in neural semantic parsers. 1 Introduction Semantic parsing is the problem of translating hu- man language into computer language, and there- fore is at the heart of natural language understand- ing. A typical semantic parsing task is question answering against a database, which is accom- plished by translating questions into executable logical forms (i.e., programs) that output their an- swers. Recent work has shown that recurrent neu- ral networks can be used for semantic parsing by encoding the question then predicting each token of the logical form in sequence (Jia and Liang, 2016; Dong and Lapata, 2016). These approaches, while effective, have two major limitations. First, they treat the logical form as an unstructured se- quence, thereby ignoring type constraints on well- formed programs. Second, they do not address en- tity linking, which is a critical subproblem of se- mantic parsing (Yih et al., 2015). This paper introduces a novel neural semantic parsing model that addresses these limitations of prior work. Our parser uses an encoder-decoder architecture with two key innovations. First, the decoder generates from a grammar that guarantees that generated logical forms are well-typed. This grammar is automatically induced from typed log- ical forms, and does not require any manual engi- neering to produce. Second, the encoder incorpo- rates an entity linking and embedding module that enables it to learn to identify which question spans should be linked to entities. Finally, we also intro- duce a new approach for training neural semantic parsers from question-answer supervision. We evaluate our parser on WIKITABLEQUES- TIONS, a challenging data set for question answer- ing against semi-structured Wikipedia tables (Pa- supat and Liang, 2015). This data set has a broad variety of entities and relations across different ta- bles, along with complex questions that necessi- tate long logical forms. On this data set, our parser achieves a question answering accuracy of 43.3% and an ensemble of 5 parsers achieves 45.9%, both of …

Content truncated to 500 words