基于深度学习的录音文本分类方法  

Method with recording text classification based on deep learning

作　　者：张彦楠 黄小红[1] 马严[1] 丛群 ZHANG Yan-nan;HUANG Xiao-hong;MA Yan;CONG Qun(Information Network Center,Beijing University of Posts and Telecommunications,Beijing 100876,China;Beijing Wrdtech Limited Company,Beijing 100876,China)

机构地区：[1]北京邮电大学信息网络中心,北京100876 [2]北京网瑞达科技有限公司,北京100876

出　　处：《浙江大学学报（工学版）》2020年第7期1264-1271,共8页Journal of Zhejiang University：Engineering Science

基　　金：中央高校基本科研专项资金资助项目(2018RC53);国家CNGI专项资助项目(CNGI-12-03-001)。

摘　　要：为了提高具有关联工单数据的录音文本的分类精确率,根据录音文本及关联数据的特点,设计基于深度学习的录音文本分类方法.针对录音文本,通过双向词嵌入语言模型(ELMo)获得录音文本及工单信息的向量化表示,基于获取的词向量,利用卷积神经网络(CNN)挖掘句子局部特征;使用CNN分别挖掘工单标题和工单的描述信息,将CNN输出的特征进行加权拼接后,输入双向门限循环单元(GRU),捕捉句子上下文语义特征;引入注意力机制,对GRU隐藏层的输出状态赋予不同的权重.实验结果表明,与已有算法相比,该分类方法的收敛速度快,具有更高的准确率.A classification method based on deep learning was designed according to the characteristics of recording text and correlation data in order to improve the classification precision of the recording text with associated work order data. The embedding of the recording text and work order information was obtained through the bidirectional word embedding language model(ELMo). Local features of the sentence were mined by using convolutional neural networks(CNN) based on the word embedding. Title and description information of the work order were separately mined by using CNN. Features extracted by CNN were concatenated with a weighting factor. Then weighted features were entered into bidirectional gated recurrent unit(GRU) in order to capture the semantic features of the context.The attention mechanism was introduced to assign different weights to the output state of the GRU hidden layer. The experimental results show that the classification method has faster convergence rate and higher accuracy compared with the existing algorithms.

关 键 词：词向量 卷积神经网络(CNN) 双向门限循环单元 注意力 文本分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]