Recognition of implicit discourse relations in paragraphing based on multigraph convolutional networks

The task of Implicit Discourse Relation Recognition (IDRR) involves detecting implicit relations between two discourse units without the use of connectives and classifying them. Currently, most models primarily adopt learning-based approaches. Some of these methods start from sentence text information, while others extensively consider the contextual relationships within discourse and external knowledge for comparing global and local information. However, these methods have not adequately placed discourse relations in a broader context and closely linked them to the textual information of sentences. Therefore, in this paper, at the paragraph level of discourse units, we propose a graph convolutional model that combines a multi-head self-attention mechanism to integrate paragraphized representations of discourse units in dependency relationships. This model obtains target embeddings from paragraphized word embeddings through BiLSTM and is experimentally validated on the PDTB2.0 dataset. Experimental results demonstrate that this model outperforms existing state-of-the-art systems on the PDTB benchmark corpus.