基于TF-IDF的卷积神经网络新闻文本分类优化  

Convolutional Neural Network for News Text Classification Optimization Based on TF-IDF

作　　者：张波 黄晓芳[1] ZHANG Bo;HUANG Xiaofang(School of Computer Science and Technology,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)

机构地区：[1]西南科技大学计算机科学与技术学院,四川绵阳621010

出　　处：《西南科技大学学报》2020年第1期64-69,共6页Journal of Southwest University of Science and Technology

基　　金：青年科学基金项目(61702429)。

摘　　要：针对使用词语级别的预训练嵌入向量初始化卷积神经网络的嵌入层在计算资源有限时存在内存溢出和训练时间长的问题,对新闻文本作出假设:去除部分不重要的词语不会影响最终分类效果,并基于TF-IDF提出一种类别关键词提取方法。通过提取类别的关键词减少词表,进一步减小嵌入矩阵的大小。在THUCNews数据集上进行的实验表明:当嵌入矩阵参数减少近89%时,在CPU的训练时间减少约49%,模型大小减少约87%,分类性能不受影响。The training of Convolutional Neural Network(CNN)with embedded layer initialized by pre-trained words vectors is characterized by problems of Out of Memory(OOM)and longtime cost.Keen on addressing these problems,a reduction of some extraneous words without compromising the model's capability and a novel method based on TF-IDF are proposed.Key words are extracted to reduce word dictionary consequently reducing the size of embedded matrix.Experiments conducted on THUCNews dataset show when the parameters of embedded matrix are reduced by nearly 89%,the training time on the CPU is reduced by about 49%while the model size is reduced by about 87%with no adverse effect on performance of classification.

关 键 词：卷积神经网络 嵌入向量 文本分类 TF-IDF 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]