Abstract: Compared with the traditional supervised Chinese relation extraction, the method based on distant supervision can greatly avoid the shortage of training corpus, so it has received extensive attention. However, the performance of the methods based on distant supervision is seriously constrained by the wrong labels introduced in the process of constructing corpus. Therefore, in order to alleviate the impact of noisy data, a relation extraction model based on dual attention mechanism is proposed in this paper. The model can obtain the context semantic information of training instances by bidirectional gated recurrent unit network, and focus on the important semantic features in the instances through the character-level attention mechanism. At the same time, the instance-level attention mechanism is introduced to calculate the correlation between instance and the corresponding relation in multiple instances in order to reduce the weight of noisy data. The experimental results on the Chinese character relationship corpus based on hudong encyclopedia show that the model compared to the single attention mechanism  models can effectively utilize the semantic information contained in the instances and reduce the influence of the wrong label instance, and get higher accuracy.

Key words: Chinese relation extraction, distant supervision, dual attention mechanism, Bidirectional Gated Recurrent Unit（BI-GRU）, hudong encyclopedia

摘要： 相比于传统有监督的中文关系抽取方法，基于远程监督的方法可极大地避免训练语料匮乏的问题，因此得到了广泛关注。然而，远程监督方法的性能却严重受困于构建语料过程中引入的错误标签，因此为缓解噪声数据所带来的影响，提出一种基于双重注意力机制的关系抽取模型。该模型可通过双向门限循环单元（Bidirectional Gated Recurrent Unit，BI-GRU）网络获取训练实例的双向上下文语义信息，并利用字符级注意力机制关注实例中重要的语义特征，同时在多个实例间引入实例级注意力机制计算实例与对应关系的相关性，以降低噪声数据的权重。在基于互动百科构建的中文人物关系抽取语料上的实验结果表明，该模型相比于单注意力机制模型可有效利用实例中所包含的语义信息并降低错误标签实例的影响，获取更高的准确率。

关键词: 中文关系抽取, 远程监督, 双重注意力机制, 双向门限循环单元（BI-GRU）, 互动百科