基于卷积神经网络和关键词策略的实体关系抽取方法  

Entity Relation Extraction Based on Convolutional Neural Network and Keywords Strategy

作　　者：王林玉[1] 王莉[1] 郑婷一[1,2] 

机构地区：[1]太原理工大学计算机科学与技术学院,晋中030600 [2]山西能源学院电气与动力工程系,晋中030600

出　　处：《模式识别与人工智能》2017年第5期465-472,共8页Pattern Recognition and Artificial Intelligence

基　　金：国家高技术研究发展计划(863计划)(No.2014AA015204);山西省自然科学基金项目(No.2014011022-1);中国科学院计算技术研究所网络数据科学与技术重点实验室课题资助~~

摘　　要：针对传统的实体关系抽取方法在定义特征过程中费时且容易造成错误传播,及现有深度学习方法依靠单一词向量学习特征的不足,文中提出基于卷积神经网络和关键词策略的实体关系抽取方法.在原始词向量的基础上,通过基于句子级的关键词抽取算法(TP-ISP)获得类别关键词特征.类别关键词的加入提高类别区分度,同时弥补网络自动学习特征的不足.在网络训练阶段,采用分段最大池化策略,减少传统最大池化策略的信息丢失.实验表明,文中方法有利于提升实体关系抽取结果.The conventional relation extraction methods are time consuming, the error propagation in feature selection is likely to emerge, and the deep learning methods only depend on word embeddings to learn features. Aiming at these problems, a relation extraction method based on convolutional neural network and keywords strategy is proposed. Based on feature of the word embeddings, the keywords feature is acquired by the term proportion-inverse sentence proportion （TP-ISP） keywords extraction algorithm based on sentence. Thus, the category division is increased and the deficiency of the network to automatically learn features from sentence is remedied. In the network training process, the chunk-based max pooling strategy is adopted to reduce the information loss by the traditional max-over-time pooling strategy. The experiment demonstrates that the proposed method improves the results of entity relation extraction.

关 键 词：卷积神经网络(CNN) 关系抽取 关键词特征 分段最大池化策略 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.1[自动化与计算机技术—控制科学与工程]