A Label Dependence-Aware Sequence Generation Model for Multi-Level Implicit Discourse Relation Recognition

Implicit discourse relation recognition (IDRR) is a challenging but crucial task in analysis. Most existing methods train multiple models to predict multi-level labels independently, while ignoring the dependence between hierarchically structured labels. In this paper, we consider IDRR as conditional label sequence generation and propose Label Dependence-aware Sequence Generation Model (LDSGM) for it. Specifically, first design attentive encoder learn global representation of an input instance its level-specific contexts, where integrated obtain better embeddings. Then, employ decoder output predicted top-down manner, higher-level are directly used guide prediction at current level. We further develop mutual learning enhanced training method exploit bottom-up direction, which captured by auxiliary introduced during training. Experimental results on PDTB dataset show that our model achieves state-of-the-art performance IDRR. release code https://github.com/nlpersECJTU/LDSGM.