CLFM: Contrastive Learning and Filter-attention Mechanism for Joint Relation Extraction

Relation extraction is a fundamental task in natural language processing, which involves extracting structured information from textual data. Despite the success of joint methods in recent years, most of them still have the propagation of cascade errors. Specifically, the error in former step will be accumulated into the final combined triples. Meanwhile, these methods also encounter another challenges related to insufficient interaction between subtasks. To alleviate these issues, this paper proposes a novel joint relation extraction model that integrates a contrastive learning approach and a filter-attention mechanism. The proposed model incorporates a potential relation decoder that utilizes contrastive learning to reduce error propagation and enhance the accuracy of relation classification, particularly in scenarios involving multiple relationships. It also includes a relation-specific sequence tagging decoder that employs a filter-attention mechanism to highlight more informative features, alongside an auxiliary matrix that amalgamates information related to entity pairs. Extensive experiments are conducted on two public datasets and the results demonstrate that this approach outperforms other models with the same structure in recall and F1. Moreover, experiments show that both the contrastive learning strategy and the proposed filter-attention mechanism work well.