A machine reading comprehension model with counterfactual contrastive learning for emotion-cause pair extraction

The emotion-cause pair extraction (ECPE) task aims to extract emotional clauses and corresponding cause clauses from documents, and this task is increasingly being recognized for its wide range of applications. In recent years, some studies have proposed using machine reading comprehension (MRC) to solve this problem and have demonstrated their effectiveness. However, these studies have not proposed methods to guide models learning the relationship between queries and documents. This study proposes an MRC model for emotion-cause extraction based on counterfactual contrastive learning. We model the relationship between queries and clauses and use contrastive learning to guide the model to learn this relationship explicitly. A novel approach is proposed to enhance the efficacy of contrastive learning and acquire more comprehensive semantic information from queries. Specifically, counterfactual methods are employed to invert the emotional polarity of emotional clauses within cause queries, which are regarded as negative instances in contrastive learning. Additionally, we incorporate emotional polarity into queries to improve the accuracy of pairing emotional clauses and cause clauses. The experimental results on publicly available Chinese and English datasets demonstrate that our model achieves state-of-the-art performance in ECPE tasks.