A Syntactic Neural Model for General-Purpose Code Generation

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 440–450 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1041 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 440–450 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1041 A Syntactic Neural Model for General-Purpose Code Generation Pengcheng Yin Language Technologies Institute Carnegie Mellon University pcyin@cs.cmu.edu Graham Neubig Language Technologies Institute Carnegie Mellon University gneubig@cs.cmu.edu Abstract We consider the problem of parsing natu- ral language descriptions into source code written in a general-purpose programming language like Python. Existing data- driven methods treat this problem as a lan- guage generation task without considering the underlying syntax of the target pro- gramming language. Informed by previ- ous work in semantic parsing, in this pa- per we propose a novel neural architecture powered by a grammar model to explicitly capture the target syntax as prior knowl- edge. Experiments find this an effective way to scale up to generation of complex programs from natural language descrip- tions, achieving state-of-the-art results that well outperform previous code generation and semantic parsing approaches. 1 Introduction Every programmer has experienced the situation where they know what they want to do, but do not have the ability to turn it into a concrete im- plementation. For example, a Python programmer may want to “sort my list in descending order,” but not be able to come up with the proper syn- tax sorted(my list, reverse=True) to real- ize his intention. To resolve this impasse, it is common for programmers to search the web in natural language (NL), find an answer, and mod- ify it into the desired form (Brandt et al., 2009, 2010). However, this is time-consuming, and thus the software engineering literature is ripe with methods to directly generate code from NL descriptions, mostly with hand-engineered meth- ods highly tailored to specific programming lan- guages (Balzer, 1985; Little and Miller, 2009; Gvero and Kuncak, 2015). In parallel, the NLP community has developed methods for data-driven semantic parsing, which attempt to map NL to structured logical forms ex- ecutable by computers. These logical forms can be general-purpose meaning representations (Clark and Curran, 2007; Banarescu et al., 2013), for- malisms for querying knowledge bases (Tang and Mooney, 2001; Zettlemoyer and Collins, 2005; Berant et al., 2013) and instructions for robots or personal assistants (Artzi and Zettlemoyer, 2013; Quirk et al., 2015; Misra et al., 2015), among oth- ers. While these methods have the advantage of being learnable from data, compared to the pro- gramming languages (PLs) in use by program- mers, the domain-specific languages targeted by these works have a schema and syntax that is rela- tively simple. Recently, Ling et al. (2016) have proposed a data-driven code generation method for high-level, general-purpose PLs like Python and Java. This work treats code generation as a sequence-to- sequence modeling problem, and introduce meth- ods to generate words from character-level mod- els, and copy variable names …

Content truncated to 500 words