面向中文基础教育知识图谱的关系抽取模型

A relation extraction model oriented to the knowledge graph of Chinese basic education

在构建中文基础教育知识图谱过程中,使用远程监督的方法能够有效解决训练语料匮乏的问题,同时使用神经网络模型能够提升构建过程中关系抽取的准确率.为了缓解远程监督中引入的错误标签带来的影响,模型通过双向门限循环单元(bidirectional gated recurrent unit)获取双向上下文中的语义信息,同时引入句子层注意力机制,动态降低噪声数据的权重.在基于中文信息技术教材教辅和百度百科的基础上构建的知识库上的关系抽取实验表明,引入句子层注意力机制能够促进模型的关系抽取效果,模型的准确率相比于中文主流关系抽取方法提高了4%~5%,能更好地应用于知识图谱的构建.

In the process of constructing a knowledge graph of Chinese basic education,the use of remote supervision can effectively solve the problem of lack of training corpora,and the use of neural network models can improve the efficiency of relation extraction during the construction process.In order to mitigate the impact of incorrect labels introduced in remote supervision,the model obtains semantic information in a bidirectional context through a bidirectional gated recurrent unit,while introducing a sentence-level attention mechanism to dynamically reduce the weight of noisy data.The relationship extraction experiments on the knowledge base built on the basis of Chinese information technology teaching materials and baidu encyclopedia show that the introduction of sentence-level attention mechanism can promote the relationship extraction effect of the model.The mainstream relation extraction method is improved by 4%-5%,which can be better applied to the construction of knowledge graph.