Reproduce “SEntFiN 1.0: Entity-Aware Sentiment Analysis for Financial News” on OpenResearch

About this paper

Ankur Sinha, Satishwar Kedas, Rishu Kumar, Pekka Malo

Researchers introduce SEntFiN 1.0, a human-annotated financial news dataset designed to capture entity-specific sentiments in headlines with multiple or conflicting entities. Their entity-aware sentiment analysis framework, leveraging fine-tuned RoBERTa models, achieves an average accuracy of 94.29% across sentiment classes, demonstrating that these extracted sentiments hold predictive value for aggregate market movements.

Key insights

• Domain-specific pre-trained language models like finBERT and RoBERTa significantly outperform other approaches in entity-aware financial sentiment extraction, achieving average accuracy increments exceeding 7%.
• Lexicon-based approaches, when enhanced with specialized domain knowledge and n-gram features, can achieve competitive performance, highlighting the continued importance of feature engineering.
• Fine-grained, entity-aware sentiments extracted from news flow possess demonstrable economic value, with after-market sentiments statistically predicting next-day market opening prices over long durations.

Summary by alphaXiv.