Yang Suizhu, Liu Yanxia, Jiang Yuantong, Liu Zhiqiang
School of Software Engineering, South China University of Technology, China.
School of Software Engineering, South China University of Technology, China.
Neural Netw. 2023 Jan;157:193-201. doi: 10.1016/j.neunet.2022.10.008. Epub 2022 Oct 19.
Distant supervision (DS) can automatically generate annotated data for relation extraction (RE) with knowledge bases and corpora. The existing DS methods that train on bags selected by attention mechanism are susceptible to noisy bags and neglect useful information in noisy bags. In this paper, we propose DCSR, a novel DS method which utilizes deep clustering to obtain refined superbag representations for solving the wrong labeling problem. we substitute deep clustering for selective attention to construct superbags, capturing helpful information between spatially-close bags, including noisy bags. Moreover, we implement data augmentation on the input sentences to handle the long-tail problem. Experiments on the NYT2010 and NYT-H datasets show that our method can effectively improve RE and significantly outperforms state-of-the-art methods.
远程监督（DS）可以利用知识库和语料库自动生成用于关系抽取（RE）的标注数据。现有的基于注意力机制选择的包进行训练的DS方法容易受到噪声包的影响，并且会忽略噪声包中的有用信息。在本文中，我们提出了DCSR，这是一种新颖的DS方法，它利用深度聚类来获得精炼的超包表示，以解决错误标注问题。我们用深度聚类代替选择性注意力来构建超包，捕捉空间上接近的包（包括噪声包）之间的有用信息。此外，我们对输入句子进行数据增强以处理长尾问题。在NYT2010和NYT-H数据集上的实验表明，我们的方法可以有效地改进关系抽取，并且显著优于现有最先进的方法。
