【摘要】 关系抽取是信息抽取的主要任务之一,远程监督作为关系抽取中的一种有效的方法,已成功地应用于包含上千关系的大型语料库.然而,远程监督造成的错误标注会影响关系抽取的性能.为了缓解这一问题,现有的远程监督关系抽取方法选择每个实体对中一个最好的句子或通过注意力机制赋予每个句子不同的权重.但这些方法并不能完全解决错误标注的问题.本文提出了一种新的方法来寻找错误标注或简单的实例,并通过动态改变损失函数的方式来降低它们在批量梯度下降中的权重.在NYT-Freebase公共数据集上的实验结果表明,本文提出的方法优于基线方法,能够有效提高远程监督关系抽取的准确率.更多还原

【Abstract】 Relation extraction is an important task in information extraction. Distantsupervision for relation extraction is an efficient method,and it has been successfully applied to large corpus w ith thousands of relations. How ever,the w rong labeling problem w ill hurt the performance of relation extraction. To alleviate this issue,most of the recent existing distant supervision methods get instances by selecting one best sentence or calculating attention w eights over the bag of sentences. These methods are not optimal,so the instances still exist problems. In this paper,w e propose a novel method to find the instances that might be noise or simple,and reduce their w eights in M ini-Batch Gradient Descent by changing the loss function dynamically. Experiments show that our method outperforms the baseline methods on a w idely used dataset.更多还原