计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (18): 176-188.DOI: 10.3778/j.issn.1002-8331.2311-0068

• 模式识别与人工智能 • 上一篇    下一篇

因果关系表示增强的跨领域命名实体识别

刘小明,曹梦远,杨关,刘杰,王杭   

  1. 1.中原工学院 计算机学院, 郑州 450007  
    2.国家语委中国语言智能研究中心,北京 100089
    3.郑州市文本处理与图像理解重点实验室,郑州 450007
    4.河南省网络舆情监测与智能分析重点实验室,郑州 450007
    5.北方工业大学 信息学院,北京 100144
  • 出版日期:2024-09-15 发布日期:2024-09-13

Causal Relationship Representation Enhanced Cross-Domain Named Entity Recognition

LIU Xiaoming, CAO Mengyuan, YANG Guan, LIU Jie, WANG Hang   

  1. 1.School of Computer Science, Zhongyuan University of Technology, Zhengzhou 450007, China
    2.Research Center for Language Intelligence of China, Beijing 100089, China
    3.Zhengzhou Key Laboratory of Text Processing and Image Understanding, Zhengzhou 450007, China
    4. Henan Key Laboratory on Public Opinion Intelligent Analysis, Zhengzhou 450007, China
    5.School of Information Science, North China University of Technology, Beijing 100144, China
  • Online:2024-09-15 Published:2024-09-13

摘要: 跨领域命名实体识别在现实应用中,尤其在目标领域数据稀缺的小样本场景中具有重要价值。然而,现有方法主要是通过特征表示或模型参数共享实现的跨领域实体能力迁移,未充分考虑由于样本选择偏差而引起的虚假相关性问题。为了解决跨领域中的虚假相关性问题,提出一种因果关系表示增强的跨领域命名实体识别模型,将源域的语义特征表示与目标域的语义特征表示进行融合,生成一种增强的上下文语义特征表示。通过结构因果模型捕捉增强后的特征变量与标签之间的因果关系。在目标域中应用因果干预和反事实推断策略,提取存在的直接因果效应,从而进一步缓解特征与标签之间的虚假相关性问题。该方法在公共数据集上进行了实验,实验结果得到了显著提高。

关键词: 跨领域命名实体识别, 迁移学习, 因果关系, 结构因果模型, 语义特征表示

Abstract: Cross-domain named entity recognition is of great value in real-world applications, especially in few-shot scenarios with scarce data in the target domain. However, the existing methods are mainly cross-domain entity capability migration achieved by feature representation or model parameter sharing, which does not fully consider the false correlation problem due to sample selection bias. To address the problem of false correlation in cross-domain, a causal representation-enhanced cross-domain named entity recognition model is proposed, in which the semantic feature representation of the source domain is first fused with that of the target domain to generate an enhanced contextual semantic feature representation. Then, the causal relationships between the enhanced feature variables and the labels are captured by a structural causal model. Finally, causal intervention and counterfactual inference strategies are applied in the target domain to extract the presence of direct causal effects, thus further mitigating the problem of spurious correlation between features and labels. The method is experimented on a public dataset and the experimental results are significantly improved.

Key words: cross-domain named entity recognition, transfer learning, causal relationship, structural causal model, semantic feature representation