关键词: 实体关系抽取, 记忆增强, 跨度筛选, 预训练语言模型, 跨句子上下文

Abstract: Entity and relation extraction (ERE) is usually processed in a pipeline manner, but this pipeline method only relies on the output of the previous task, which leads to information interaction problems between entity recognition and relation extraction, and easily causes error propagation problems. A memory enhancement model based on span screening and bidirectional dependency (MEERE) is proposed to address the above issues. This model introduces a memory-like mechanism so that each task can not only utilize the output of the previous task but also reversely affect the previous task, thereby capturing the complex interaction between entities and relations. Secondly, an entity span screening mechanism is introduced to further alleviate error propagation. This mechanism ensures that only high-quality entities are used for relation extraction by dynamically screening and verifying entity spans in the joint module, thereby improving the robustness and accuracy of the model. Finally, the table decoding method is used to deal with the relationship overlap problem well. Experimental results on three widely used benchmark datasets (ACE05, SciERC, and CoNLL04) show that the MEERE model shows significant advantages in the ERE task. Compared with existing methods, it has improved the accuracy of both entity recognition and relation extraction, especially in reducing error propagation and improving overall model stability. Compared with Tab-Seq on the CoNLL04 dataset, MEERE has significant improvements in both entity and relation extraction, with an increase of 0.4 percentage points in entity F1 and 2.7 percentage points in strict evaluation F1. Compared with PURE-F, MEERE achieves more than 10 times and better relation extraction performance. These results verify the effectiveness of the proposed memory-enhanced model in exploring the interaction between entities and relations.

Key words: Entity relation extraction, Memory enhancement, Span screening, Pre-trained language models, Cross-sentence context