计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (23): 169-177.DOI: 10.3778/j.issn.1002-8331.2105-0173

• 模式识别与人工智能 • 上一篇    下一篇

结合噪声网络的强化学习远程监督关系抽取

谢斌红,王恩慧,张英俊   

  1. 太原科技大学 计算机科学与技术学院,太原 030024
  • 出版日期:2022-12-01 发布日期:2022-12-01

Distant Supervision Relation Extraction Based on Reinforcement Learning with Noisy Network

XIE Binhong, WANG Enhui, ZHANG Yingjun   

  1. School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China
  • Online:2022-12-01 Published:2022-12-01

摘要: 针对目前远程监督关系抽取任务中存在的错误标注问题,提出使用强化学习策略设计噪声指示器,通过与由关系分类器和噪声数据组成的环境相交互,动态识别每个关系类别的假正例与假负例,并为其重新分配正确的关系标签,从而将噪声数据转换成有用的训练样本,有利于提高远程监督关系抽取模型的性能;另外,在训练过程中,通过在策略网络权重上添加噪声,平衡策略网络的探索和利用问题,从而增强噪声指示器的探索能力,使噪声指示器更准确地选择出能够正确表达实体-关系的句子。在Freebase对齐NYT公共数据集上的实验结果表明,提出的方法可以显著提高远程监督关系抽取模型的性能,表明模型拥有识别并纠正噪声数据标签的能力,可以更好地学习关系特征。

关键词: 远程监督关系抽取, 强化学习, 噪声网络, 假负例

Abstract: Aiming at the noisy labeling problem in the current distant supervision relation extraction task, this paper proposes a reinforcement learning strategy to design a noisy indicator. By interacting with the environment composed of relation classifier and noisy data, the false positive instances and false negative instances of each relation category are dynamically identified, and the correct relation labels are redistributed, thus, the noisy data is transformed into useful training samples, which is helpful to improve the performance of the distant supervision relation extraction model. In addition, in the process of training, noise is added to the weight of policy network to balance the exploration and utilization of policy network, so as to enhance the exploration ability of noisy indicator and make the noisy indicator more accurately select sentences that can correctly express entity relationship. The experimental results on freebase aligned NYT public dataset show that the proposed method can significantly improve the performance of the distant supervision relation extraction model, which shows that the model has the ability to recognize and correct noisy data labels, and can better learn the relation features.

Key words: distant supervision relation extraction, reinforcement learning, noisy network, false negative instances