Weakly-Supervised Neural Semantic Parsing with a Generative Ranker

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 356–367 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 356 Weakly-supervised Neural Semantic Parsing with a Generative Ranker Jianpeng Cheng and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB jianpeng.cheng@ed.ac.uk mlap@inf.ed.ac.uk Abstract Weakly-supervised semantic parsers are trained on utterance-denotation pairs, treating logical forms as latent. The task is chal- lenging due to the large search space and spuriousness of logical forms. In this paper we introduce a neural parser-ranker system for weakly-supervised semantic parsing. The parser generates candidate tree-structured logical forms from utterances using clues of denotations. These candidates are then ranked based on two criterion: their likelihood of executing to the correct denotation, and their agreement with the utterance semantics. We present a scheduled training procedure to balance the contribution of the two ob- jectives. Furthermore, we propose to use a neurally encoded lexicon to inject prior domain knowledge to the model. Experiments on three Freebase datasets demonstrate the effectiveness of our semantic parser, achieving results within the state-of-the-art range. 1 Introduction Semantic parsing is the task of converting natural language utterances into machine-understandable meaning representations or logical forms. The task has attracted much attention in the literature due to a wide range of applications ranging from question answering (Kwiatkowski et al., 2011; Liang et al., 2011) to relation extraction (Krishna- murthy and Mitchell, 2012), goal-oriented dialog (Wen et al., 2015), and instruction understanding (Chen and Mooney, 2011; Matuszek et al., 2012; Artzi and Zettlemoyer, 2013). In a typical semantic parsing scenario, a logical form is executed against a knowledge base to pro- duce an outcome (e.g., an answer) known as deno- tation. Conventional semantic parsers are trained on collections of utterances paired with annotated logical forms (Zelle and Mooney, 1996; Zettle- moyer and Collins, 2005; Wong and Mooney, Figure 1: Overview of the weakly-supervised neu- ral semantic parsing system. 2006; Kwiatkowksi et al., 2010). However, the la- beling of logical forms is labor-intensive and chal- lenging to elicit at a large scale. As a result, al- ternative forms of supervision have been proposed to alleviate the annotation bottleneck faced by se- mantic parsing systems. One direction is to train a semantic parser in a weakly-supervised setting based on utterance-denotation pairs (Clarke et al., 2010; Kwiatkowski et al., 2013; Krishnamurthy and Mitchell, 2012; Cai and Yates, 2013), since such data are relatively easy to obtain via crowd- sourcing (Berant et al., 2013a). However, the unavailability of logical forms in the weakly-supervised setting, renders model training more difficult. A fundamental challenge in learning semantic parsers from denotations is finding consistent logical forms, i.e., those which execute to the correct denotation. This search space can be very large, growing exponentially as compositionality increases. Moreover, consistent logical forms unavoidably introduce a certain de- gree of spuriousness — some of them will acci- dentally execute to the correct denotation without reflecting the meaning …

Content truncated to 500 words