计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (22): 65-71.DOI: 10.3778/j.issn.1002-8331.2106-0278

• 理论与研发 • 上一篇    下一篇

融合细粒度实体类型的多特征关系分类算法

左亚尧,易彪,黎文杰   

  1. 广东工业大学 计算机学院,广州 510006
  • 出版日期:2022-11-15 发布日期:2022-11-15

Multi-Feature Relationship Classification Algorithm Fused with Fine-Grained Entity Types

ZUO Yayao, YI Biao, LI Wenjie   

  1. School of Computer, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2022-11-15 Published:2022-11-15

摘要: 近年来,基于深度学习的关系分类多集中于注意力机制的改进或语义信息的优化两方面,但此类方法提取的特征往往较为单一,忽略了实体类型对关系分类的影响,且存在语义信息学习不完整等问题。提出一种新的关系分类方法Type-SBNE,针对实体类型学习任务,引入细粒度的实体类型信息,并通过对每个实体类型中的所有实体向量取平均生成实体类向量,再分别通过实体与句意信息学习获得其对应的特征向量,通过拼接融合得到复合语义特征,最后采用全连接层和Softmax函数来预测实体对之间的关系。Type-SBN基于细粒度的实体类型信息,丰富了实体的特征,有效加强了每个实体在上下文中的表达。实验表明,Type-SBNE模型可更好地完成关系分类任务,效果优于对比方法。

关键词: 关系分类, BERT, Nystr?mformer, 细粒度实体类向量

Abstract: In recent years, relationship classification based on deep learning has mostly focused on the improvement of attention mechanism or the optimization of semantic information, but the features extracted by such methods are often relatively single, ignoring the impact of entity types on relationship classification, and there is semantic information problem such as incomplete learning. This paper proposes a new relationship classification method Type-SBNE, which introduces fine-grained entity type information for entity type learning tasks, and generates entity class vectors by averaging all entity vectors in each entity type, and then the corresponding feature vectors are obtained through entity information learning and sentence meaning information learning, and composite semantic features are obtained through splicing and fusion. Finally, the fully connected layer and Softmax function are used to predict the relationship between entity pairs. Based on the fine-grained entity type information, the characteristics of the entity are enriched, and the expression of each entity in the context is effectively strengthened. Experiments show that the Type-SBNE model can better complete the relationship classification task, and the effect is improved compared with other algorithms.

Key words: relation classification, BERT, Nystr?mformer, fine-grained entity type vector