一种融合摘要与主体特征的混合神经网络文本主题分类方法  

A Hybrid Neural Network Combining Text Summary and Main Part Features for Topic Classification

作　　者：张伟智 陈羽中[1,2] 郭昆 林涵阳[3] ZHANG Weizhi;CHEN Yuzhong;GUO Kun;LIN Hanyang(College of Mathematics and Computer Sciences,Fuzhou University,Fuzhou 350116;Fujian Provincial Key Laboratory of Network Computing and Intelligent Information Processing,Fuzhou 350116;Jiangsu Start Dima Data Processing Co.,Ltd.,Kunshan 215332)

机构地区：[1]福州大学数学与计算机科学学院,福州350116 [2]福建省网络计算与智能信息处理重点实验室,福州350116 [3]江苏实达迪美数据处理有限公司,昆山215332

出　　处：《计算机与数字工程》2020年第5期1100-1107,共8页Computer & Digital Engineering

摘　　要：文本主题分类是自然语言处理中的重要任务之一,也是深度学习方法的重要应用领域。目前用于文本分类的深度学习方法多利用词-句-文本的层次结构来挖掘文本的语义及句法结构信息,以构建网络模型。然而,文本的语义和句法结构信息不仅仅包含于文本的层次结构中,限制了深度网络模型的文本主题分类能力。针对上述问题,论文提出一种融合摘要与主体特征的混合神经网络文本主题分类方法:首先分别抽取文本中的摘要和主体部分,之后根据神经网络的特性,使用卷积神经网络学习文本摘要中的关键局部特征,同时使用长短期记忆网络学习文本主体中句内及句间的上下文时序特征,最后设计融合文本摘要和主体特征的级联神经网络提升模型对文本主题的理解能力。实验结果表明,论文提出的方法能有效提升文本主题分类精度。Text topic classification is one of the important tasks of text data mining,and also an important application field of deep learning methods.At present,most deep neural networks for text make use of the hierarchical structure in which words form sentences and sentences form texts to build models so as to mine the semantic and syntactic information of texts.However,the semantic and syntactic information are not only hidden in the hierarchical structure,which limits the ability of text topic classification in deep network models.To solve the above problem,a hybrid neural network combining text summary and main part features are presented.First,the summary and main part from a text are extracted respectively,then,according to the characteristics of the different neural networks,Convolutional Neural Network is used to learn the local feature,and Long-Short Term Memory is used to learn the context feature between words within a sentence and sentences within a text simultaneously.Finally,the model combining the two features is desingned to improve the ability to understand the text topic.Experimental results show that the method can improve the classification accuracy efficiently.

关 键 词：文本主题分类 卷积神经网络 长短期记忆网络 词向量 文本摘要 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]