基于BERT-BiGRU-CNN的文本分类模型  

Text Classification Model Based on BERT-BiGRU-CNN

作　　者：隋德义 祁云嵩 SUI Deyi;QI Yunsong(School of Computer Science,Jiangsu University of Science and Technology,Zhenjiang 212000)

机构地区：[1]江苏科技大学计算机学院,镇江212000

出　　处：《计算机与数字工程》2025年第2期474-479,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61471182)资助。

摘　　要：针对卷积神经网络(CNN)存在丢失结构信息和循环神经网络(RNN)很难处理具有长期依赖的信息等问题。论文提出一种BERT-BiGRU-CNN(BeGC)模型,该模型采用预训练模型BERT作为词嵌入层,以更好地获取语义信息,通过双向门控循环单元(BiGRU)获取上下文语义特征,再进行最大池化处理得到文本的重要信息,从而解决了CNN和RNN在文本分类中的局限性问题。最后在公开数据集THUCNews上进行对比实验,结果表明,该模型的性能优于大多数分类模型,证明了模型的可行性。Convolutional neural network(CNN)has some problems,such as the loss of structural information and the difficul⁃ty of recurrent neural network(RNN)in dealing with long-term dependent information.This paper proposes a BERT-BiGRU-CNN(BeGC)model,which uses the pre-training model BERT as the text embedding layer to better obtain the semantic information of the text.The context semantic features are obtained through the bidirectional gate recurrent unit(BiGRU),and then the important information of the text is obtained through the maximum pooling processing,so as to solve the limitations of CNN and RNN in text classification.Finally,a comparative experiment is carried out on the public data set of THUCNews.The results show that the perfor⁃mance of this model is better than most classification models,which proves the feasibility of the model.

关 键 词：文本分类 语义特征 BERT BiGRU 最大池化 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]