A Novel Joint Extraction Model for Entity Relations Using Interactive Encoding and Visual Attention

Abstract

Relationship extraction is a fundamental task in natural language processing, with applications ranging from knowledge graph construction to information retrieval. Existing entity-relationship joint extraction models have made significant strides in this field. However, they still face limitations in effectively utilizing interaction information between subjects and objects, as well as capturing the spatial location relationships between entities. In this paper, we propose a novel relationship extraction model that addresses these limitations. Our model introduces innovative techniques to harness interaction information between subjects and objects. We employ subject gates, object gates, entity gates, and relationship gates to partition and filter interaction information between relationship triples during the encoding phase. Additionally, we leverage an attention mechanism inspired by the visual domain to capture spatial location relationships between entities during the decoding phase, transforming the entity-relationship joint extraction task into a table-filling task. To evaluate the effectiveness of our model, we conducted extensive experiments on multiple datasets, including WebNLG, NYT, and ADE. Our model achieved impressive F1 values of 93.65%, 92.58%, and 86.16% on these datasets, respectively, outperforming state-of-the-art models.