Connective-aware interaction attention for implicit discourse relation classification

Implicit discourse relation classification, identifying relationships between arguments without explicit linguistic cues, is a challenging task. Previous studies have shown that connectives are important for recognizing implicit discourse relations. Most previous works applied connective prediction as an auxiliary task to promote knowledge transfer from connectives to labels which did not make full use of the relational mapping information of connectives. In this work, we propose an innovative Connective-aware Interactive Attention (CAIA) joint learning approach. Specifically, we use BERT to predict connectives and incorporate connective information into the interaction of the attention mechanism. Our experimental results on the PDTB dataset show that our approach achieves competitive results compared to recent state-of-the-art systems.