| 引用本文： |
-
牙珊珊,陈定甲,郑宏春,李航,覃晓.基于关系模式与深度强化学习的DS数据去噪模型[J].广西科学院学报,2022,38(4):403-411. [点击复制]
- YA Shanshan,CHEN Dingjia,ZHENG Hongchun,LI Hang,QIN Xiao.DS Data Denoising Model Based on Relation Pattern and Deep Reinforcement Learning[J].Journal of Guangxi Academy of Sciences,2022,38(4):403-411. [点击复制]
|
|
| 摘要: |
| 远程监督(Distant Spervision，DS)数据集中存在大量错误标注的数据，而现有的DS数据集去噪方法通常只考虑针对具有标签的数据进行去噪，没有充分利用无标签数据，导致去噪效果不佳。本文提出一种新型DS数据去噪模型——Pattern Reinforcement Learning Model (PRL模型)：首先利用基于关系模式的正样例抽取算法提取DS数据集中高质量的有标签数据；然后利用Filter-net作为分类器，提取DS数据集中高质量的无标签数据；最后将高质量的有标签数据和无标签数据作为深度强化学习(Reinforcement Learning，RL)方法的训练数据集，获得去噪效果更好的远程监督数据集。将PRL模型应用于New York Times(NYT)数据集，并以去噪后的数据集来训练PCNN+ONE、CNN+ATT、PCNN+ATT 3个模型。实验结果表明，经过PRL模型对数据集进行去噪后，这些模型的性能得以提升。因此，PRL模型是一种轻量的数据去噪模型，可以提升基于深度神经网络模型的性能。 |
| 关键词: 自然语言处理 关系分类 远程监督 迁移学习 去噪方法 |
| DOI：10.13657/j.cnki.gxkxyxb.20221209.009 |
| 投稿时间：2022-07-26修订日期：2022-09-03 |
| 基金项目:国家自然科学基金项目(61962006)和广西高校中青年教师科研基础能力提升项目(2022KY0378)资助。 |
|
| DS Data Denoising Model Based on Relation Pattern and Deep Reinforcement Learning |
|
YA Shanshan1, CHEN Dingjia1, ZHENG Hongchun1, LI Hang2, QIN Xiao1
|
| (1.Guangxi Key Laboratory of Human Computer Interaction and Intelligent Decision Making, Nanning Normal University, Nanning, Guangxi, 530100, China;2.School of Artificial Intelligence, Guangxi Minzu University, Nanning, Guangxi, 530006, China) |
| Abstract: |
| There are a large number of mislabeled data in the Distant Supervision (DS) dataset.The existing denoising methods of DS dataset usually only consider denoising the data with labels,and do not make full use of the unlabeled data,resulting in poor denoising effect.In this article,a new DS data denoising model-Pattern Reinforcement Learning Model (PRL model) is proposed.Firstly,the positive sample extraction algorithm based on relational pattern is used to extract high-quality labeled data in DS dataset.Then use Filter-net as a classifier to extract high-quality unlabeled data from DS dataset.Finally,high-quality labeled data and unlabeled data are used as the training dataset of the deep Reinforcement Learning (RL) method to obtain a remote supervision dataset with better denoising effect.The PRL model is applied to the New York Times (NYT) dataset,and the denoised dataset is used to train PCNN+ONE,CNN +ATT,PCNN+ATT three models.The experimental results show that the performance of these models is improved after the dataset is denoised by PRL model.Therefore,the PRL model is a lightweight data denoising model that can improve the performance of deep neural network-based models. |
| Key words: natural language processing relation classification distant supervision strengthen learning denoising method |
