Paragraph-Level Interaction Discourse Relation Recognition Model via Neural Tensor Network

In Natural Language Processing (NLP), if separated from other parts of the paragraph, the semantics of the sentence or clause cannot be fully understood. The semantic understanding of a sentence or clause also depends on all discourse relations and the whole paragraph-level discourse structure. To better understand the semantics of sentences by improving the performance of discourse relation classification, the Paragraph-level Interaction model via Neural Tensor Network (PINTN) is constructed to model the interdependence between Discourse Unit (DU) and the continuity and pattern of discourse relations. On the basis of the discourse representation, the Neural Tensor Network (NTN) is further used to explore deeper semantic interaction. So as to better recognize discourse relations. The experimental results on PDTB dataset show that the PINTN model by constructed is effective. Its performance in implicit and explicit discourse relation recognition is superior to other models. The experimental result shows that the recognition of implicit discourse relations is affected by the recognition of contiguous explicit discourse relations.