Implicit Discourse Relation Recognition Based on Dual-filter Label-aware Contrastive Learning

Implicit discourse relation recognition is a challenging task in discourse analysis, which aims to identify the semantic logical relation between two arguments that lack connectives. To model the global hierarchy that contains all labels and ignore noise information in the discourse arguments, this paper proposes a Dual-Filter Label-aware Contrastive Learning model. First, the model uses an improved graph encoder named Graphormer to model the label hierarchy and obtain a better label representation. In addition, we proposed label-aware contrastive learning, and adopted a filter strategy combining supervised and unsupervised to learn more robust representation of discourse relations. Experimental results on open data sets show that the model achieves optimal performance on existing baselines.