基于深度学习和多模态特征融合的情感分类方法

A Deep Learning-Based Multimodal Feature Fusion Approach for Sentiment Classification

摘要: 本研究提出了一种面向股票评论情感分析的多模态深度学习模型TMV-FinSent，通过协同建模文本、图像和数值三类异质特征，实现对投资者情感倾向与市场情绪强度的精准识别。模型采用预训练基座网络和混合专家网络，有效提升了情感分析的精度和泛化能力。实验结果表明，TMV-FinSent在准确率、召回率和F1分数等指标上显著优于传统基准模型，证明了多模态融合在情感识别中的有效性。该模型为金融情绪分析提供了新的技术路径，具备良好的扩展性和应用前景，能应用于智能投研和舆情风险监测等领域。

Abstract: This study proposes TMV-FinSent, a multimodal deep learning model for sentiment analysis of stock-related comments, which achieves accurate identification of investor sentiment polarity and market emotion intensity by jointly modeling heterogeneous features from text, images, and numerical data. The model leverages pretrained foundation models and a mixture-of-experts architecture, significantly enhancing the accuracy and generalization capability of sentiment analysis. Experimental results demonstrate that TMV-FinSent substantially outperforms traditional baseline models in terms of accuracy, recall, and F1-score, validating the effectiveness of multimodal fusion in sentiment recognition. The proposed model offers a novel technical approach for financial sentiment analysis, exhibiting strong extensibility and promising applications in areas such as intelligent investment research and public opinion risk monitoring.