摘要:远监督关系抽取的最大优势是通过知识库和自然语言文本的自动对齐生成标记数据.这种简单的自动对齐机制在将人从繁重的样本标注工作中解放出来的同时，不可避免地会产生各种错误数据标记，进而影响构建高质量的关系抽取模型.针对远监督关系抽取任务中的标记噪声问题，提出"最终句子对齐的标签是基于某些未知因素所生成的带噪观测结果"这一假设.并在此假设的基础上，构建由编码层、基于噪声分布的注意力层、真实标签输出层和带噪观测层的新型关系抽取模型.模型利用自动标记的数据学习真实标签到噪声标签的转移概率，并在测试阶段，通过真实标签输出层得到最终的关系分类.随后，研究带噪观测模型与深度神经网络的结合，重点讨论基于深度神经网络编码的噪声分布注意力机制以及深度神经网络框架下不均衡样本的降噪处理.通过以上研究，进一步提升基于带噪观测远监督关系抽取模型的抽取精度和鲁棒性.最后，在公测数据集和同等参数设置下进行带噪观测远监督关系抽取模型的验证实验，通过分析样本噪声的分布情况，对在各种样本噪声分布下的带噪观测模型进行性能评价，并与现有的主流基线方法进行比较.结果显示，所提出的带噪观测模型具有更高的准确率和召回率.



Abstract:The great advantage of distant supervision relation extraction is to generate labeled data automatically through knowledge bases and natural language texts. This simple automatic alignment mechanism liberates people from heavy labeling work, but inevitably produces various incorrect labeled data meanwhile, which would have an influential effect on the construction of high-quality relation extraction models. To handle noise labels in the distant supervision relation extraction, here it is assumed that the final label of sentence is based on noisy observations generated by some unknown factors. Based on this assumption, a new relation extraction model is constructed, which consists of encoder layer, attention based on noise distribution layer, real label output layer, and noisy observation layer. In the training phase, transformation probabilities are learned from real label to noisy label by using automatically labeled data, and in the testing phase, the real label is obtained through the real label output layer. This study proposes to combine the noise observation model with deep neural network. The attention mechanism of noise distribution is focused based on deep neural network, and unbalanced samples are denoised of under the framework of deep neural network, aiming to further improve the performance of distant supervision relation extraction based on noisy observation. To examine its performance, the proposed method is applied to a public dataset. The performance of distant supervision relation extraction model is evaluated under different distribution families. The experimental results illustrate the proposed method is more effective with higher precision and recall, compared to the existing methods.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5929