Emotion-cause pair extraction based on machine reading comprehension model

In this paper, we propose a BERT-based framework for Emotion-Cause Pair Extraction (ECPE) task. Given a passage, the ECPE task aims to jointly extract (1) emotion-related clauses and (2) cause clauses (the clause caused the emotion). Our framework is featured by the following two novel designs. First, we formulate the emotion and cause extraction task as a machine reading comprehension (MRC) task. The MRC task is to read a given text passage, and then answer questions by comprehending the article. In our formulation, we treat the ECPE passage as MRC input and pose questions like (Which clauses cause the emotions?). The idea is to leverage the power of MRC model based on recent pre-trained language model. Second, we formulate the emotion-cause pair detection as contextual relatedness detection problem, which can be also effectively addressed by pre-trained language model. The experiment results based on benchmarking datasets demonstrate the effectiveness of the proposed approach; we advance the state-of-the-art results from 61% to 65% in terms of F1 scores.