基于VDCNN与LSTM混合模型的中文文本分类研究  

Research on Chinese Text Classification Based on Hybrid Model of VDCNN and LSTM

作　　者：彭玉青 宋初柏 闫倩 赵晓松 魏铭 PENG Yuqing;SONG Chubai;YAN Qian;ZHAO Xiaosong;WEI Ming(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401

出　　处：《计算机工程》2018年第11期190-196,共7页Computer Engineering

基　　金：河北省自然科学基金重点项目(F2016202144);河北省自然科学基金面上项目(F2017202145)

摘　　要：自然语言在结构上存在一定的前后依赖性,且将中文文本直接转化为向量时会使维度过高,从而导致现有文本分类方法精度较低。为此,建立一种超深卷积神经网络(VDCNN)与长短期记忆网络(LSTM)相结合的混合模型。通过VDCNN的深度结构来提取文本向量的特征,利用LSTM具有存储历史信息的特点提取长文本的上下文依赖关系,同时引入词嵌入将文本转换为低维度向量。在Sogou语料库和复旦大学中文语料库上进行实验,结果表明,相对CNN+rand、LSTM等模型,该混合模型可以有效提升文本分类的精确率。Due to the structural dependence of natural language,and the high dimensionality when Chinese text is directly transformed into vector,the accuracy of existing text classification methods is low.To solve this problem,a hybrid model of Very Deep Convolution Neural Network(VDCNN)and Long Short-Term Memory network(LSTM)is proposed.The depth structure of VDCNN is used to extract the features of text vectors,the context dependence of long text is extracted by using LSTM’s feature of storing historical information,and word embedding is introduced to transform text into low-dimensional vector.Experimental on Sogou corpus and Fudan University Chinese corpus show that,the hybrid model can effectively improve the accuracy of text classification compared with CNN+rand and LSTM.

关 键 词：文本分类 卷积神经网络 长短期记忆网络 词嵌入 深度残差网络 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]