Evaluating semantic insensitivity in financial sentiment analysis: implications for sustainable financial decision-making
Evaluating Semantic Insensitivity in Financial Sentiment Analysis: Implications for Sustainable Financial Decision-Making
Researchers propose an adversarial attack framework called Agent-driven Semantic Inversion (ASI), which attacks financial sentiment analysis models through domain-specific semantic manipulation and validates its effectiveness on the widely deployed FinBERT. ASI uses an LLM-driven multi-agent pipeline to identify qualifying financial news sentences and apply targeted lexical substitutions, successfully deceiving FinBERT in 90.85% of qualifying cases, with the model often maintaining its original sentiment prediction even when financial performance indicators are fundamentally reversed. Qualitative analysis and token-level investigation show that FinBERT relies heavily on contextual keywords and syntactic patterns rather than genuine logical understanding of financial semantics. The authors emphasize the need for logic-aware, more robust financial NLP models to ensure the reliability of automated decision-making systems such as automated trading, portfolio management, and risk assessment.