Scholarlink

<p>Financial sentiment analysis is crucial for understanding market trends and making informed investment decisions. Traditional models often rely on unimodal data, limiting their effectiveness in capturing the full context of financial news. This study proposes the Multimodal Uncertainty-Aware Financial Sentiment Analysis (MUFSAN) framework, which integrates textual, visual, and acoustic data to enhance sentiment predictions. The methodology involves preprocessing each modality using advanced models-LLaMA for text, Vision Transformers for visual data, and Wav2Vec 2.0 for audio. Features from these modalities are integrated using concatenation, dimensionality reduction via PCA, and multi-head attention mechanisms. Bayesian inference with Hamiltonian Monte Carlo quantifies uncertainty in the predictions. Experimental evaluation on the FMSA-SC dataset demonstrates the framework’s superior performance, achieving an accuracy of 85.23%, precision of 84.01%, recall of 83.54%, and F1-score of 83.77%, outperforming all baseline models. The MUFSAN framework also provides well-calibrated uncertainty estimates, with a Predictive Entropy of 0.1754 and Expected Calibration Error of 0.0456. Sensitivity and ablation studies confirm the robustness and significance of each component. This comprehensive approach not only improves prediction accuracy but also offers reliable uncertainty quantification, making it a valuable tool for financial sentiment analysis and decision-making.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A transformer-based uncertainty quantification framework for multimodal financial sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • E. Subha,
  • V. Jothi Prakash,
  • S. Arul Antran Vijay

摘要

Financial sentiment analysis is crucial for understanding market trends and making informed investment decisions. Traditional models often rely on unimodal data, limiting their effectiveness in capturing the full context of financial news. This study proposes the Multimodal Uncertainty-Aware Financial Sentiment Analysis (MUFSAN) framework, which integrates textual, visual, and acoustic data to enhance sentiment predictions. The methodology involves preprocessing each modality using advanced models-LLaMA for text, Vision Transformers for visual data, and Wav2Vec 2.0 for audio. Features from these modalities are integrated using concatenation, dimensionality reduction via PCA, and multi-head attention mechanisms. Bayesian inference with Hamiltonian Monte Carlo quantifies uncertainty in the predictions. Experimental evaluation on the FMSA-SC dataset demonstrates the framework’s superior performance, achieving an accuracy of 85.23%, precision of 84.01%, recall of 83.54%, and F1-score of 83.77%, outperforming all baseline models. The MUFSAN framework also provides well-calibrated uncertainty estimates, with a Predictive Entropy of 0.1754 and Expected Calibration Error of 0.0456. Sensitivity and ablation studies confirm the robustness and significance of each component. This comprehensive approach not only improves prediction accuracy but also offers reliable uncertainty quantification, making it a valuable tool for financial sentiment analysis and decision-making.