基于权重词向量与改进TextCNN的中文新闻分类  

Chinese News Classification Based on Weighted Word Vector and Improved TextCNN

作　　者：万铮 王芳 黄树成 WAN Zheng;WANG Fang;HUANG Shucheng(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212114,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212114

出　　处：《软件导刊》2023年第9期59-64,共6页Software Guide

基　　金：国家自然科学基金项目(61772244)。

摘　　要：为了解决单一深度学习模型提取信息不充分、分类效果不佳的问题,提出一种混合多神经网络的BA-Info-CNN-BiLSTM模型。该模型使用BERT作为词嵌入层,获得词的向量表示,再使用注意力机制让词获得不同权重;然后一方面将其送入改进的文本卷积神经网络(InfoCNN)中获取文本的局部信息特征,另一方面送入双向的长短时记忆网络(Bi-LSTM)中获得文本的全局信息特征;最后将提取到的局部信息和全局信息进行拼接融合,送入softmax函数中进行分类,得到分类结果。通过与其他模型进行对比实验,该模型获得了较好的分类效果,在新浪新闻和搜狐新闻数据集上分别取得了95.07%和84.95%的准确率,在一定程度上解决了单一模型捕获信息不充分的问题。In order to solve the problem of insufficient information extraction and poor classification effect of a single deep learning model,this paper proposes a BA-InfoCNN-BiLSTM model with a hybrid multi-neural network.The model uses BERT as the word embedding layer to obtain the vector representation of the word,and then uses the attention mechanism to obtain different weights for the word,and then sends it to the improved text convolutional neural network on the one hand to obtain The local information features of the text,on the other hand,are sent to the bidirectional long short-term memory network to obtain the global information features of the text,and finally the extracted local in-formation and global information.After the features are spliced and fused,they are sent to the softmax function for classification,and the clas-sification results are obtained.After comparing experiments with other models,this model has achieved good classification results.It has achieved 95.07%and 84.95%accuracy on the Sina news data set and Sohu news data set respectively,which solves the problem of insufficient information captured by single model to a certain extent.

关 键 词：文本分类 词嵌入 注意力机制 卷积神经网络 循环神经网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]