Combining BERT with Bi-LSTM for Emotion-Cause Pair Extraction

Abstract

The emotion-cause pair extraction aims to extract emotion clause and the underlying cause clause for the emotion from documents. Existing methods for extracting emotion-cause pairs mostly focus on semantic representations of clauses, ignoring emotion and cause features in the emotional causality, which weakens the representation learning ability of emotion-cause pairs. In this paper, we propose an Emotion-Cause Pair Feature Extraction model (ECPFE) to solve the problem. Pre-training model BERT and graph attention network are used to obtain clause semantic information. Emotion extractor and cause extractor based on Bi-LSTM are utilized to learn valuable features from semantic information. The emotion extractor captures emotion features and cause extractor captures cause features in clauses. And these valuable features are used to design the specific emotion-cause pair representations to improve the model’s representation learning ability. Moreover, relative position information between clauses obtained by functional relative position encoding is embedded in the specific representations to enhance position-awareness of the model. Experimental results demonstrate that our ECPFE model outperforms the existing models by 3.4% in F1 score.