Entity-Aware Biaffine Attention for Constituent Parsing

Abstract

Constituency parsing is the process of analyzing a sentence by breaking it down into sub-phrases also known as constituents. Although many deep neural models have achieved state-of-the-art results on this task, few consider entity-violating issue, i.e. an entity cannot form a complete sub-tree in the resultant constituent parsing tree. To attack this issue, this paper proposes an entity-aware biaffine attention model for constituent parsing. It leverages entity information for a potential phrase when conducting biaffine attention between the start and end words of the phrase. In the absence of the proper metric for comparison, the entity violating rate (EVR) as a new metric is introduced here to evaluate how many the final parsing trees suffer from entity violating issue. The lower the EVR, the better the model. This metric from a brand perspective helps us understand the potential of existing arts. Experiments on three publicly popular datasets including ONTONOTES, PTB and CTB show that our model achieves the lowest EVR while almost achieving the same performance in terms of the three conventional metrics, i.e., precision, recall, and F1-score. Moreover, extensive experiments of sentence sentiment analysis as a downstream application further exhibit the efficacy of our model and the validity of the proposed metric EVR.