Multi-View Cooperative Learning with Invariant Rationale for Document-Level Relation Extraction

Abstract

Document-level relation extraction (RE) is a complex and significant natural language processing task, as the massive entity pairs exist in the document and are across sentences in reality. However, the existing relation extraction methods (deep learning) often use single-view information (e.g., entity-level or sentence-level) to learn the relational information but ignore the multi-view information, and the explanations of deep learning are difficult to be reflected, although it achieves good results. To extract high-quality relational information from the document and improve the explanations of the model, we propose a multi-view cooperative learning with invariant rationale (MCLIR) framework. Firstly, we design the multi-view cooperative learning to find latent relational information from the various views. Secondly, we utilize invariant rationale to encourage the model to focus on crucial information, which can empower the performance and explanations of the model. We conduct the experiment on two public datasets, and the results of the experiment demonstrate the effectiveness of MCLIR.