PGLR: Pseudo Graph and Label Reuse for Entity Relation Extraction

Named entity recognition and relation extraction are essential for structured information construction. However, existing methods, particularly in the field of relation prediction, struggled with information redundancy. In addition, conventional classification models only focus on predicting labels and comparing them to the ground truth, without considering the values and semantics of the labels. In this paper, we propose a novel end-to-end joint entity recognition and relation extraction model named PGLR (Pseudo Graph and Label Reuse for Entity Relation Extraction), which mainly decomposes the task into three hierarchical components to cope with the above issues. Firstly, the model generates a fully connected pseudo-graph that connects all possible entities and relations, followed by a confidence-based module to remove nodes and edges associated with non-relational facts. Secondly, a label reuse approach is employed to aggregate predicted labels and external knowledge, and the representations are reconstructed using an optimal re-parameterization technique. Thirdly, a gating mechanism is utilized to gauge the confidence of the reconstructed representations. The experimental results from three benchmark datasets (ACE04, ACE05, and SciERC) indicate that the proposed model achieves greater accuracy, faster inference speed, and requires fewer parameters.