结合注意力机制与残差网络的远程监督关系抽取

Distant Supervision for Relation Extraction Based on Combining Attention and Residual Network

关系抽取是信息提取领域的重要任务之一。虽然有监督的学习方法在关系抽取任务中取得了相当好的效果,但监督样本的缺乏限制了有监督学习方法的应用,于是有人提出将远程监督思想用于关系抽取任务中。现今,远程监督的关系抽取被广泛应用于从文本中发现新的关系实体。然而远程监督不可避免地伴随着噪音数据,显然这些噪音数据将会影响关系抽取任务的效果。为了尽可能地解决这个问题,论文结合Special Self Attention和Deep Residual Learning(ResNet)设计了一种新的深度学习模型(ARCNN),期望可以动态地减少噪音数据的影响,更好地提取文本的深层信息。实验结果显示论文的模型可以有效地减轻远程监督数据的噪音影响,并更好地抽取出实体对的关系。

Relation Extraction(RE)is an important task of Information Extraction.Although supervised learning methods have achieved quite good results in relational extraction tasks,the lack of supervised samples limits the application of supervised learning methods.So some people proposed to use the remote supervision idea in the relationship extraction task.Up to now,distant supervised RE has been widely used to find novel relational facts from text.But distant supervision inevitably accompanies with the wrong labelling problem,and absolutely these noisy data will generate the negative effects on RE task.In order to alleviate this issue,this paper designs a novel deep learning model(ARCNN)with Deep residu-al learning(ResNet),which is expected to dynamically reduce the impact of those noisy in-stances,and could extract more deep information of text.Experimental results show that this model can effectively reduce the influence of noisy data,and improve the performance for relation extraction.