作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (8): 63-68. doi: 10.19678/j.issn.1000-3428.0065258

• 人工智能与模式识别 • 上一篇    下一篇

融合混合表征的小样本关系抽取方法

刘昊鑫1, 董超1, 勾智楠2, 高凯1,*   

  1. 1. 河北科技大学 信息科学与工程学院, 石家庄 050018
    2. 河北经贸大学 信息技术学院, 石家庄 050061
  • 收稿日期:2022-07-15 出版日期:2023-08-15 发布日期:2023-08-15
  • 通讯作者: 高凯
  • 作者简介:

    刘昊鑫(1997—),男,硕士研究生,主研方向为自然语言处理

    董超,讲师、博士

    勾智楠,讲师、博士

  • 基金资助:
    河北省自然科学基金(F2022208006); 河北省高等学校科学技术研究项目(QN2020198)

Few-Shot Relation Extraction Method Fusing with Hybrid Representation

Haoxin LIU1, Chao DONG1, Zhinan GOU2, Kai GAO1,*   

  1. 1. School of Information Science and Engineering, Hebei University of Science and Technology, Shijiazhuang 050018, China
    2. School of Information Technology, Hebei University of Economics and Business, Shijiazhuang 050061, China
  • Received:2022-07-15 Online:2023-08-15 Published:2023-08-15
  • Contact: Kai GAO

摘要:

针对传统原型网络无法有效解决噪声数据带来的类原型偏差的问题,提出一种融合混合表征的小样本关系抽取方法,通过引入关系表述和实体提及作为辅助信息,并通过动态构建关系类别原型表示来提升模型对噪声数据的处理能力,进而获得更精确的原型判别性表示。首先,使用BERT预训练语言模型对文本进行特征提取,利用实体提及提取实体关系表示,构建局部原型,并设计关系注意力模块对关系表述和支持实例进行语义计算和关键特征选择;然后,提出基于混合表征的动态原型网络构建模块,分别利用局部原型和关系注意力模块动态构建实体表征类原型和关系表征类原型;最后,融合两者形成更具判别性的混合表征原型点,进一步增强原型的关系表示能力。实验结果表明,在公开数据集FewRel 1.0上,融合混合表征的小样本关系抽取方法相较于基线模型,在不同子任务设置下均取得了较高的准确率,最高可达96.26%,验证了所提方法能有效利用辅助信息缓解类原型偏差问题,获得较好的关系抽取效果。

关键词: 关系抽取, 小样本学习, 原型网络, 注意力机制, 混合表征

Abstract:

Addressing the problem in which the traditional prototypical network cannot effectively solve the prototype deviation caused by noisy data, this study proposes a few-shot relation extraction method fusing with hybrid representation.This method introduces relation representation and entity mention as auxiliary information, and the prototype points of the relation categories are dynamically constructed to improve the processing ability of the model for noisy data. Therefore, a more accurate prototype discriminant representation is obtained. First, the BERT pre-trained language model extracts textual features.Simultaneously, entity mentions are used to extract entity relation representations, establishing local prototypes. A relation attention module is proposed to perform semantic computation and key feature selection on relation representations and support instances.Second, a dynamic prototypical network building module is proposed based on hybrid representation, where the local prototype and the relation attention module are used to dynamically construct entity representation prototype points and relation representation prototype points, respectively. Finally, the two prototype points are fused to form a more discriminative hybrid representation prototype point, which further enhances the relation representation ability of the prototype.The experimental results show that on the public dataset FewRel 1.0, the few-shot relation extraction method fusing with hybrid representation achieves up to 96.26% higher accuracy than the baseline model under different subtask settings.These results validate the efficacy of the proposed method in utilizing auxiliary information to alleviate the problem of prototype deviation and obtaining a better relationship extraction effect.

Key words: relation extraction, few-shot learning, prototypical network, attention mechanism, hybrid representation