摘要
针对传统文本分类模型存在中文文本语义特征的提取能力有限等问题,提出一种BERT-WWM融合双通道语义特征的文本情感分析模型,学习更深层次的文本语义特征表示.首先采用BERT-WWM获取中文文本的动态特征向量表示并传入双通道卷积神经网络-双向长短期记忆网络(CNN-BiSLTM)中进行特征提取,然后将BiLSTM通道提取文本的上下文特征向量经注意力(Attention)层动态权重调整后,与CNN通道提取的文本局部语义特征向量进行融合,以增强模型的文本特征提取能力.最后,将融合特征经过全连接层和Softmax函数得出文本的情感倾向.实验结果表明,相比传统的单通道模型和多通道混合模型,所提模型在准确率Acc和综合评价指标F1上分别提高了0.91%和1.40%,证明了该模型在中文文本情感分析任务中的有效性和可行性.
Aiming at the problem of the limited extraction ability of Chinese text semantic features in traditional text classification models,a BERT-WWM text emotion analysis model is proposed to learn the deeper representation of text semantic features.Firstly,BERT-WWM is used to obtain the dynamic feature vector representation of Chinese text and it is passed into the two-channel convolutional neural network(CNN-BiSLTM)for feature extraction.Then,the context feature vector of text extracted by BiLSTM channel is adjusted by the dynamic weight of the Attention layer.The model is fused with the local semantic feature vector extracted by CNN channel to enhance the capability of text feature extraction.Finally,the fusion features are passed through the full connection layer and Softmax function to obtain the emotional tendency of the text.The experimental results show that compared with the traditional single-channel model and multi-channel mixed model,the accuracy of Acc and the comprehensive evaluation index F1 are improved by 0.91%and 1.40%,respectively,which proves the effectiveness and feasibility of the model in Chinese text sentiment analysis.
作者
赵雪峰
狄恒西
柏长泽
仲兆满
Zhao Xuefeng;Di Hengxi;Bai Changze;Zhong Zhaoman(School of Computer Engineering,Jiangsu Ocean University,Lianyungang 222005,China)
出处
《南京师范大学学报(工程技术版)》
2025年第4期18-27,48,共11页
Journal of Nanjing Normal University(Engineering and Technology Edition)
基金
国家自然科学基金项目(72174079)
江苏省“青蓝工程”优秀教学团队项目(2022-29)。
关键词
文本情感分析
BERT-WWM
双通道语义特征
卷积神经网络
双向长短期神经网络
text sentiment analysis
BERT-WWM
dual-channel semantic feature
convolutional neural network
bidirectional long and short term memory network
