首页|期刊导航|现代信息科技|融合层级语义与多尺度卷积的短文本情感分析模型

融合层级语义与多尺度卷积的短文本情感分析模型OA

A Short Text Sentiment Analysis Model Integrating Hierarchical Semantics and Multi-Scale Convolution

中文摘要英文摘要

针对垂直领域短文本情感分析中语义碎片化及特征利用不足问题,文章提出融合层级语义与多尺度卷积(Multi-Scale Convolution,MSC)的分类模型 MSBERT.该模型基于轻量化 Transformer 结构,通过层级注意力机制对各编码层进行自适应权重融合,并结合多尺度一维卷积提取多粒度局部特征,实现全局与局部语义协同建模.在11 987条外卖评论数据集上的实验显示,模型二分类F1值达0.883 0,优于SVM、LSTM及BERT基准模型.结果证明,该方法在小规模数据场景下具有较强的情感识别与泛化能力.

To address the problems of fragmented semantics and insufficient feature utilization in sentiment analysis of short texts in vertical domains,this paper proposes MSBERT,a classification model that integrates hierarchical semantics and Multi-Scale Convolution(MSC).Built on a lightweight Transformer architecture,the model performs adaptive weighted fusion of the encoding layers through a hierarchical attention mechanism,combines multi-scale 1D convolution to extract multi-granularity local features,and achieves collaborative modeling of global and local semantics.Experimental results on a dataset of 11 987 takeaway reviews show that the model achieves an F1-score of 0.883 0 in binary classification,outperforming SVM,LSTM,and BERT baseline models.The results demonstrate that the proposed method has strong sentiment recognition and generalization capability in small-scale data scenarios.

周洁

浙江万里学院 物流与电子商务学院,浙江 宁波 315100

信息技术与安全科学

短文本情感分析BERT层级注意力多尺度卷积特征融合评论数据

short text sentiment analysisBERThierarchical attentionMulti-Scale Convolutionfeature fusionreview data

《现代信息科技》 2026 (18)

80-85,92,7

10.19850/j.cnki.2096-4706.2026.18.014

评论