Generating computer programs from natural language descriptions

Abstract

This thesis addresses the problem of learning to translate natural language into preexisting programming languages supported by widely-deployed computer systems. Generating programs for existing computer systems enables us to take advantage of two important capabilities of these systems: computing the semantic equivalence between programs, and executing the programs to obtain a result. We present probabilistic models and inference algorithms which integrate these capabilities into the learning process. We use these to build systems that learn to generate programs from natural language in three different computing domains: text processing, solving math problems, and performing robotic tasks in a virtual world. In all cases the resulting systems provide significant performance gains over strong baselines which do not exploit the underlying system capabilities to help interpret the text.