Interaction and Fusion of Rich Textual Information Network for Document-level Relation Extraction

Detecting relations between entities across multiple sentences in a document, referred to as document-level relation extraction, poses a challenge in natural language processing. Graph networks have gained widespread application for their ability to capture long-range contextual dependencies in documents. However, previous studies have often been limited to using only two to three types of nodes to construct document graphs. This leads to insufficient utilization of the rich information within the documents and inadequate aggregation of contextual information. Additionally, relevant relationship labels often co-occur in documents, yet existing methods rarely model the dependencies of relationship labels. In this paper, we propose the Interaction and Fusion of Rich Textual Information Network (IFRTIN) that simultaneously considers multiple types of nodes. First, we utilize the structural, syntactic, and discourse information in the document to construct a document graph, capturing global dependency relationships. Next, we design a regularizer to encourage the model to capture dependencies of relationship labels. Furthermore, we design an Adaptive Encouraging Loss, which encourages well-classified instances to contribute more to the overall loss, thereby enhancing the effectiveness of the model. Experimental results demonstrate that our approach achieves a significant improvement on three document-level relation extraction datasets. Specifically, IFRTIN outperforms existing models by achieving an F1 score improvement of 0.67% on Dataset DocRED, 1.2% on Dataset CDR, and 1.3% on Dataset GDA. These results highlight the effectiveness of our approach in leveraging rich textual information and modeling label dependencies for document-level relation extraction.