Sequence-based Structured Prediction for Semantic Parsing

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1341–1350, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Sequence-based Structured Prediction for Semantic Parsing Chunyang Xiao and Marc Dymetman Xerox Research Centre Europe 6-8 Chemin de Maupertuis Meylan, 38240, France chunyang.xiao, marc.dymetman @xrce.xerox.com Claire Gardent CNRS, LORIA, UMR 7503 Vandoeuvre-l`es-Nancy, F-54500, France claire.gardent@loria.fr Abstract We propose an approach for semantic parsing that uses a recurrent neural net- work to map a natural language question into a logical form representation of a KB query. Building on recent work by (Wang et al., 2015), the interpretable log- ical forms, which are structured objects obeying certain constraints, are enumer- ated by an underlying grammar and are paired with their canonical realizations. In order to use sequence prediction, we need to sequentialize these logical forms. We compare three sequentializations: a direct linearization of the logical form, a linearization of the associated canonical realization, and a sequence consisting of derivation steps relative to the underlying grammar. We also show how grammati- cal constraints on the derivation sequence can easily be integrated inside the RNN- based sequential predictor. Our experi- ments show important improvements over previous results for the same dataset, and also demonstrate the advantage of incor- porating the grammatical constraints. 1 Introduction Learning to map natural language utterances (NL) to logical forms (LF), a process known as seman- tic parsing, has received a lot of attention recently, in particular in the context of building Question- Answering systems (Kwiatkowski et al., 2013; Berant et al., 2013; Berant and Liang, 2014). In this paper, we focus on such a task where the NL question may be semantically complex, leading to a logical form query with a fair amount of compo- sitionality, in a spirit close to (Pasupat and Liang, 2015). Given the recently shown effectiveness of RNNs (Recurrent Neural Networks), in particu- lar Long Short Term Memory (LSTM) networks (Hochreiter and Schmidhuber, 1997), for perform- ing sequence prediction in NLP applications such as machine translation (Sutskever et al., 2014) and natural language generation (Wen et al., 2015), we try to exploit similar techniques for our task. However we observe that, contrary to those appli- cations which try to predict intrinsically sequen- tial objects (texts), our task involves producing a structured object, namely a logical form that is tree-like by nature and also has to respect cer- tain a priori constraints in order to be interpretable against the knowledge base. In our case, building on the work “Building a Semantic Parser Overnight” (Wang et al., 2015), which we will refer to as SPO, the LFs are gener- ated by a grammar which is known a priori, and it is this grammar that makes explicit the structural constraints that have to be satisfied by the LFs. The SPO grammar, along with generating logi- cal forms, generates so-called “canonical forms” (CF), which are direct textual realizations of the LF that, although they are not “natural” English, transparently convey the meaning of the LF …

Content truncated to 500 words