Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (20): 138-146.DOI: 10.3778/j.issn.1002-8331.2206-0459

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Joint Entity Relation Extraction Based on Two-Stage Decoding

CHANG Sijie, LIN Haotian, JIANG Jing   

  1. Internet of Things and Robotics Lab, Smart City College, Beijing Union University, Beijing 100101, China
  • Online:2023-10-15 Published:2023-10-15

融合双阶段解码的实体关系联合抽取方法

常思杰,林浩田,江静   

  1. 北京联合大学 智慧城市学院 物联网与机器人实验室,北京 100101

Abstract: The concatenated decoding method has solved some overlapping problems by directly optimizing the triples in the existing joint entity relation extraction task. However, this method will cause the imbalance problem of entity recognition to the decoded entities under a specific relation. Although the order problem of triples has been solved only by use of the set prediction method to synchronous decode entities and relations, it also leads to the problems of weak connection among different entities and poor interaction between entities and relations. This paper proposes a joint extraction model of entity relations that integrates two-stage decoding in order to further improve the effect of the joint extraction model, which including entity decoding in the cascade decoding stage and relationship decoding in the set prediction network stage. The model in this paper is divided into three parts:Firstly, Bert is used for encoding, which pays attention to the context information effectively. Secondly, the recognition of entities uses a cascade decoding strategy, which can extract the entity information without relationship constraints, and adequate identification of entities. Finally, the triples fused with entity information are embedded into the ensemble prediction network to decode the entity-relation triplet and to strengthen the connection between entity and relation. Based on the open data sets of the New York Times(NYT), WebNLG and ACE2005, the experimental results show that the model proposed in this paper is basically superior to the baseline model, which verifies the effectiveness of the model.

Key words: entity relation joint extraction, overlapping problems, cascade decoding, ensemble prediction

摘要: 在现有的实体关系联合抽取任务中,级联解码的方法直接对三元组进行优化,解决了一部分重叠问题,但是在特定关系下解码的实体,造成实体识别不平衡问题。仅用集合预测的方法可以同时解码出实体和关系,虽然解决了三元组的顺序问题,但也导致实体之间联系性不强、实体和关系之间交互性差的问题。为了进一步提高联合抽取模型的效果,提出一种融合双阶段解码的实体关系联合抽取模型,包括级联策略下的实体解码与集合预测网络阶段的关系解码。该模型分为三个部分:采用Bert进行编码,有效关注到了上下文的信息;采用级联解码的策略先对实体识别,得到不受关系限制的实体信息,充分识别实体;将融合了实体信息的表示嵌入集合预测网络解码出实体-关系三元组,加强实体与关系的联系。在公开数据集纽约时报(The New York Times,NYT)、WebNLG和ACE2005上的实验结果表明,所提出的模型基本优于基线模型,验证了该模型的有效性。

关键词: 实体关系联合抽取, 重叠问题, 级联解码, 集合预测