Dual Hierarchical Contrastive Learning for Multi-level Implicit Discourse Relation Recognition

Jing Xu, Ruifang He, Haodong Zhao, Huijie Wang, Lei Zeng

Abstract: Implicit discourse relation recognition (IDRR) is a challenging but vital task in discourse analysis, which aims at classifying logical relations between arguments. Previous work infuses ACCL takes external knowledge or label semantics to alleviate data scarcity, which either brings noise or underutilizes semantic information contained in label embedding. Meanwhile, it is difficult to model label hierarchy. In this paper, we make full use of label embedding as positives and negatives for our dual hierarchical contrastive learning framework, which contains two parts: 1) hierarchical label contrastive loss (HLCL), which promotes fine-grained labels to be more similar to correlative medium-grained labels than related coarse-grained labels. 2) arguments and connectives contrastive loss (ACCL), which makes arguments aggregate around correlative fine-grained labels. The two modules interact with each other, making the similarity between arguments and correlative fine-grained labels are higher than that with related coarse-grained labels. In this process, the multi-level label semantics are integrated to arguments, which provides guidance for classification. Experimental results show that our method achieves competitive performance against state-of-the-art system.