Mention Distance-aware Interactive Attention with Multi-step Reasoning for document-level relation extraction

Fu Zhang, Jiapeng Wang, Huangming Xu, Honglin Wu, Jingwei Cheng, Weijun Li

Abstract: Document-level relation extraction (DocRE) is widely used in various natural language processing and artificial intelligence applications, aiming to identify relations between entities in a document that may appear multiple times or span multiple sentences. Previous methods have relied on global document features to generate a unified representation for each entity, which may result in the loss of entity-specific information and fail to capture the nuanced relations between different entity pairs. In this paper, we propose the Mention Distance-aware Interactive Attention with Multi-step Reasoning (MDIAMR) model, which leverages attention mechanisms and graph neural networks to address these challenges. Specifically, we introduce an entity interaction attention module to explicitly model the distance information between mentions and dynamically aggregate all mentions of each entity. This approach highlights the importance of different mentions to the entity, thereby enhancing the entity’s representation. Additionally, we design a multi-step reasoning module based on graph attention networks (GAT), which iteratively updates entity pair representations by considering information from neighboring pairs. We also incorporate a cosine similarity loss function to help the model differentiate between entity pairs that are related or unrelated to the target entity pair. Experimental results demonstrate that our model outperforms competitive baseline models on two publicly available datasets.