A Joint Entity and Relation Extraction Approach Using Dilated Convolution and Context Fusion

Abstract

In recent years, researchers have shown increasing interest in joint entity and relation extraction. However, existing approaches overlook the interaction between words at different distances and the significance of contextual information between entities. We believe that the correlation strength of word pairs should be considered, and it is necessary to integrate contextual information into entities to learn better entity-level representations. In this paper, we treat named entity recognition as a multi-class classification of word pairs. We employ self-attention mechanism and design both local and multi-grained dilated convolution layers to capture spatial correlations between words. In the relation extraction module, we leverage attention from the self-attention layer to fuse localized context information into entity-pair to produce context-enhanced entity-level representations. In addition, we integrate named entity recognition and relation extraction through a multi-task learning framework, effectively leveraging the interaction between two subtasks. To validate the performance of our model, we conducted extensive experiments on joint entity and relation extraction benchmark datasets CoNLL04, ADE and SciERC. The experimental results indicate that our proposed model can achieve significant improvements over existing methods on these datasets.