计算机工程

• •    

基于图卷积神经网络的中文实体关系联合抽取

  

  • 发布日期:2020-12-21

Joint extraction of Chinese entity relation based on graph convolutional neural network

  • Published:2020-12-21

摘要: 由于中文表达方式灵活多样、句法语法复杂多变,具有更复杂的实体关系结构,现有的实体关系 联合抽取方法没有充分考虑中文句子中关系的结构特征。为此,本文提出基于图卷积神经网络的中文实体 关系联合抽取方法。该方法在双向长短时记忆网络抽取序列特征的基础上,利用图卷积神经网络编码依存 分析中的语法结构信息,借鉴新的实体标注策略的思路,构建端到端的中文实体关系联合抽取模型。实验 结果表明,该方法的 F 值可达 61.4%,相比 LSTM-LSTM 模型提高了 4.1%。实验证实图卷积神经网络能 有效编码文本先验词间关系,并有效提升实体关系抽取的性能。

Abstract: Due to the flexibility and variety of Chinese expressions and the complexity of Chinese syntax and grammar, it has a more complex entity relationship structure, and the existing entity relationship joint extraction methods do not fully consider the structural characteristics of the relationship in Chinese sentences. For this reason, this paper proposes a joint extraction method of Chinese entity and relations, which is based on graph convolutional neural network. Based on the sequence features extracted by the bidirectional long short term memory network, this method uses graph convolutional neural network to encode the grammatical structure information in dependency analysis, learns from the idea of a novel tagging strategy to label the entity, and builds an end-to-end Chinese entity relation joint extraction model. Experimental results show that this method’s F score can reach 61.4%, which is 4.1% higher than the model that extracts only sequence features. Experiments have shown that graph convolutional neural network can effectively encode the prior relation between words contained in the text, and effectively improve the performance of entity relation extraction.