An Extension of FinEntity: Benchmarking Entity-Level Sentiment for Financial Texts

An Extension of FinEntity: Enhanced Entity-Level Sentiment Classification and Commodity Case Study
By Niranjan Vijaya Krishnan, Lily Weaver, Chaeyoung Lee

📄 Read the full paper here

---

🔍 Overview

This project builds upon FinEntity (Tang et al., 2023), a framework for entity-level sentiment classification in financial news. We extend this work in two major directions:

1. Model Benchmarking: Evaluate recent language models (DeBERTa, RoBERTa, GPT-4o, LLaMA, Qwen) for financial NER + sentiment classification on the FinEntity dataset—testing with and without a CRF (Conditional Random Field) layer.
2. Commodity Case Study: Replicate the original cryptocurrency use case and expand it to commodities (Oil, Gold, Copper, Silver), analyzing correlations between entity-level sentiment and price trends.

---

📊 Key Results

Model	Positive F1	Negative F1	Micro Avg F1	Notes
DeBERTa-CRF	0.94	0.88	0.89	Best overall performance
FinBERT-CRF	0.84	0.88	0.84	Best baseline from original work
GPT-4o (fine-tuned)	0.81	0.79	0.85	Strong LLM result (fine-tuned)
GPT-3.5 (zero-shot)	0.39	0.58	0.59	Poor zero-shot performance
LLaMA / Qwen	0.76–0.79	0.66–0.70	0.70–0.73	Struggles with fine-tuning

🟢 CRF layers consistently improve performance for open-source models.
📈 In commodity analysis, entity-level sentiment generally shows stronger or equal correlation with prices (MIC) compared to sequence-level sentiment.

---

📚 Citation

If you use this repo or build on it, please cite:

@article{krishnan2024finentityextension,
  title={An Extension of FinEntity: Entity-level Sentiment Classification for Financial Texts},
  author={Krishnan, Niranjan Vijaya and Weaver, Lily and Lee, Chaeyoung},
  journal={GitHub},
  year={2024},
  url={https://github.com/niruvk/FinEntity-Extension}
}