Learning from errors: Using vector-based compositional semantics for parse reranking

Phong Le, Willem H. Zuidema, Remko Scha

Abstract: In this paper, we address the problem of how to use semantics to improve syntactic parsing, by using a hybrid reranking method: a k-best list generated by a symbolic parser is reranked based on parsecorrectness scores given by a compositional, connectionist classifier. This classifier uses a recursive neural network to construct vector representations for phrases in a candidate parse tree in order to classify it as syntactically correct or not. Tested on the WSJ23, our method achieved a statistically significant improvement of 0.20% on F-score (2% error reduction) and 0.95% on exact match, compared with the state-ofthe-art Berkeley parser. This result shows that vector-based compositional semantics can be usefully applied in syntactic parsing, and demonstrates the benefits of combining the symbolic and connectionist approaches.