Two-stage sentiment analysis with built-in error correction improves financial predictions
This paper introduces SILC, a two-stage framework that improves financial sentiment analysis by combining LLM fine-tuning with self-correction mechanisms for better accuracy.
-----
https://arxiv.org/abs/2412.19140v1
🤔 Original Problem:
Financial sentiment analysis lacks large entity-level datasets and struggles with multi-entity texts where different entities have varying sentiments within the same context.
-----
🔧 Solution in this Paper:
→ SILC uses a two-stage approach where first stage fine-tunes base LLMs (LlaMA2-7b for English, Baichuan2-7b for Chinese) to generate initial sentiment predictions
→ Second stage implements a GNN-based example retriever to find relevant correction examples from training data
→ The system filters and retains incorrectly predicted samples while sampling correct predictions to train the correction model
→ A Graph Attention Network processes both linguistic and sentiment features to retrieve similar examples for correction
-----
💡 Key Insights:
→ Three in-context examples provide optimal performance for the model
→ Retaining 60-80% of correct samples yields best correction results
→ Entity-level sentiment shows higher correlation with crypto prices than sequence-level analysis
-----
📊 Results:
→ Improved F1 score by 5.1% over previous methods on FinEntity dataset
→ Achieved RMSE of 0.07936 in Bitcoin price prediction
→ Outperformed GPT-4 and GPT-3.5 on both English and Chinese datasets
