| 谢修娟*,刘雪娟**.融合情感簇的混合神经网络短文本情感分类模型[J].高技术通讯(中文),2025,35(10):1069~1077 |
| 融合情感簇的混合神经网络短文本情感分类模型 |
| Hybrid neural network short text sentiment classification model integrating sentiment clusters |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2025. 10. 004 |
| 中文关键词: 文本情感分类； 情感簇； 胶囊网络； 双向长短期记忆网络； 注意力机制 |
| 英文关键词: text sentiment classification, sentiment clustering, capsule network, bidirectional long short-term memory, attention mechanism |
| 基金项目: |
| 作者 | 单位 | | 谢修娟* | (*东南大学成贤学院电子与计算机工程学院南京 210088)
(**南京财经大学会计学院南京 210023) | | 刘雪娟** | |
|
| 摘要点击次数: 19 |
| 全文下载次数: 33 |
| 中文摘要: |
| 针对文本分类的深度学习主流模型中存在的特征提取不全面、位置结构信息缺失等问题，提出一种融合情感簇的混合神经网络短文本情感分类模型（sentiment clustering and fusion of multiple neural networks，SCMN）。该方法首先通过双向变换器模型(bidirectional encoder representations from Transformers，BERT)预训练模型生成词向量，并进行情感簇聚类和情感权重增强；然后使用带有注意力机制的双向长短期记忆网络（bidirectional long short term memory，BiLSTM），捕获文本的上下文特征；再通过胶囊网络(capsual network，CapsNet)提取带有句子结构信息的局部语义特征并完成分类。基于公开数据集和自爬取数据集，将本文模型与深度学习主流分类模型进行对比实验及不同组件的消融实验。实验结果表明，相较于其他方法，本文模型精确率实现了平均5.5%的增长，证实了不同组件能为模型带来有效增益，提升文本情感分类效果。 |
| 英文摘要: |
| Aiming at the problem of incomplete feature extraction and lack of position structure information in the mainstream model of deep learning of text classification, a short text sentiment classification model based on sentiment clustering and fusion of multiple neural networks (SCMN) is proposed.This method first generates word vectors through bidirectional encoder representations from Transformers (BERT) pretraining models, and performs sentiment clustering and sentiment weight enhancement; then it uses the bidirectional long short-term memory (BiLSTM) network with attention mechanism to capture the context features of the text, and uses the capsual network (CapsNet) network to extract the local semantic features with sentence structure information and complete the classification. Based on publicly available datasets and self crawled datasets, the model in this paper is compared with the mainstream classification model of deep learning, and the ablation experiments of different components are conducted. The results show that compared with other methods, the accuracy of the model in this paper has achieved an average increase of 5.5%. It is confirmed that different components can bring effective gains to the model and improve the effect of text emotion classification. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
