Scholarlink

<p>Financial Sentiment Analysis has become an essential tool for guiding investment strategies and forecasting market movements. With the rapid advancement of Natural Language Processing, large language models have achieved strong performance across many tasks. However, due to the highly technical nature and domain-specific characteristics of financial texts, general-purpose models underperform on financial tasks. To address this challenge, this study builds upon the open-source pretrained model DeepSeek-R1-Distill-Qwen-7B, and enhances it through the construction of a Chain-of-Thought—annotated augmented dataset, parameter-efficient fine-tuning, and a context-aware dynamic prompting strategy. These techniques collectively lead to a 38.87% improvement in accuracy over the original model on financial sentiment analysis. In addition, prompt engineering is employed to achieve structured output and interpretable sentiment reasoning. Experimental results show that the proposed method outperforms multiple baseline models and demonstrates particularly strong zero-shot generalization across datasets, achieving higher classification accuracy than larger-scale models. These findings suggest that the proposed approach is both effective and promising for advancing financial sentiment analysis.</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

FinSent-DistillQ: A distilled large language model with chain-of-thought fine-tuning for financial sentiment analysis

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Yuan Huang,
  • Tongjia Ma,
  • Kuohai Yang,
  • Zhaoliang Zhang

摘要

Financial Sentiment Analysis has become an essential tool for guiding investment strategies and forecasting market movements. With the rapid advancement of Natural Language Processing, large language models have achieved strong performance across many tasks. However, due to the highly technical nature and domain-specific characteristics of financial texts, general-purpose models underperform on financial tasks. To address this challenge, this study builds upon the open-source pretrained model DeepSeek-R1-Distill-Qwen-7B, and enhances it through the construction of a Chain-of-Thought—annotated augmented dataset, parameter-efficient fine-tuning, and a context-aware dynamic prompting strategy. These techniques collectively lead to a 38.87% improvement in accuracy over the original model on financial sentiment analysis. In addition, prompt engineering is employed to achieve structured output and interpretable sentiment reasoning. Experimental results show that the proposed method outperforms multiple baseline models and demonstrates particularly strong zero-shot generalization across datasets, achieving higher classification accuracy than larger-scale models. These findings suggest that the proposed approach is both effective and promising for advancing financial sentiment analysis.