NS-BERT: An Explainable Neuro-Symbolic Framework for Financial Sentiment Analysis

Financial sentiment analysis remains challenging due to the complexity of natural language, specialized financial terminology and the need for interpretable predictions in high-stakes applications. Although FinBERT provides strong contextual language understanding, its predictions do not explicitly represent the reasoning underlying its outputs, making the interpretation of individual decisions challenging. This study proposes NS-BERT, a neuro-symbolic framework that augments FinBERT with a rule-based refinement mechanism designed to improve classification performance while providing traceable reasoning. The framework extracts domain-specific evidence from financial text and applies predefined symbolic rules to identify cases warranting prediction refinement. It selectively modifies FinBERT predictions when the extracted evidence satisfies applicable rule conditions and retains the original predictions otherwise. This process generates reasoning traces linking the initial prediction, extracted evidence, applied rules, and final prediction, enabling transparent examination of classification decisions. Experiments on the 50% annotator-agreement subset of the Financial PhraseBank dataset demonstrated improvements in accuracy and Macro-F1 over the FinBERT baseline. These results suggest that integrating explicit symbolic reasoning with neural contextual representations can improve financial sentiment classification while enhancing the traceability of prediction refinements.