Unifying emotion-oriented and cause-oriented predictions for emotion-cause pair extraction.

ABSTRACT

Emotion-cause pair extraction (ECPE) is an extraction task aiming to simultaneously identify the emotions and causes from the text without emotion annotations. Let ci and cj represent the emotion clause and the cause clause of a document, respectively, and we can predict one from the other and vice versa. Previous works fail to take advantage of this bidirectional opportunity. We refer to the prediction from ci to cj, i.e., ci→cj, as an emotion-oriented cause prediction (EoCP) task and the prediction from cj to ci, i.e., cj→ci, as a cause-oriented emotion prediction (CoEP) task. After redefining the ECPE task, we propose a novel unified architecture for ECPE, which incorporates EoCP and CoEP as cells and unifies them into a single-chain architecture. Additionally, we redefine emotion-cause pair extraction as a closed-loop structure detection problem to alleviate the mismatch between emotion and cause clauses. To enhance the training of the architecture, we provide a procedure for estimating the confidence of the extraction system for its emotion-cause pairs. We demonstrate the superiority of our proposed model through extensive experiments on two public datasets, achieving a new state-of-the-art performance. Furthermore, our method particularly achieves significant improvements in multiple emotion-cause pair extraction.