Knowledge Enhancement and Feature Purification for Single-stage Joint Entity and Relation Extraction

Joint entity and relation extraction aim to achieve named entity recognition and relation extraction in unstructured text. We use the form of triples (subject, relation, object) to describe entity and relation. Joint entity and relation extraction play an important role in knowledge graph construction, question and answering, data analysis, and other natural language processing domains. Most of the existing works suffered from insufficient interaction between entity features and relation features due to the extraction order. Moreover, there is a prevalence of heterogeneous representations of entities and relations. They not only increase the amount of model design and the complexity of training, but also lead to the problem of exposure bias due to the extraction order. Therefore, in this paper, we propose KFRel, where we enhance the entity and relation representations by encoding text and relation. Then, the features are fused to enhance the entity and relation representations, and adopt a feature purification module, where the features are enhanced by a feature purification module that removes feature information irrelevant to the joint entity and relation extraction while retaining feature information of relevance. In addition, we adopt a single-stage joint entity and relation extraction module to address the issue of overlapping triples. This module aims to help improve the effectiveness of joint entity and relation extraction. Comprehensive experiments are conducted on two widely-used datasets, and the experimental results demonstrate that the proposed method is effective and outperforms the state-of-the-art baselines.