Tasks like code generation and semantic parsing require mapping unstructured (or partially structured) inputs to well-formed, executable outputs. We introduce abstract syntax networks, a modeling framework for these problems. The outputs are represented as abstract syntax trees (ASTs) and constructed by a decoder with a dynamically-determined modular structure paralleling the structure of the output tree. On the benchmark Hearthstone dataset for code generation, our model obtains 79.2 BLEU and 22.7% exact match accuracy, compared to previous state-of-the-art values of 67.1 and 6.1%. Furthermore, we perform competitively on the Atis, Jobs, and Geo semantic parsing datasets with no task-specific engineering.
@article{arxiv.1704.07535,
title = {Abstract Syntax Networks for Code Generation and Semantic Parsing},
author = {Maxim Rabinovich and Mitchell Stern and Dan Klein},
journal= {arXiv preprint arXiv:1704.07535},
year = {2017}
}