Semantic Parsing with Syntax Graph of Logical Forms

Semantic parsing aims to convert natural language queries to logical forms, which are strictly structured. Recently neural semantic parsers have paid attention to structure information of target logical forms and set constraints on generating rules. In this work, we propose to use syntax graphs of both query and logical form and to utilize graph neural networks (GNN) in encoder combined with BERT pre-training and decoder with copy mechanism. Besides, we present a predicate review loss function to help GNNs in the decoder capture the syntax graph structure more precisely. Results of experiments on three datasets show that our model outperforms the baseline on MSParS, achieves state-of-art accuracy on ATIS, and has competitive performance on Job.