Shapley in Context: Explaining Financial Language with Domain Expertise
创建于 更新于
摘要
This paper investigates Shapley value-based attributions for financial text data produced by large language models, proposing domain knowledge-inspired axioms and proving that baseline Shapley values preserve these axioms to yield faithful explanations [page::0][page::1]. It provides theoretical and empirical validation across sentiment analysis and risk-factor extraction, showing that Shapley-based explanations align with financial reasoning and provide meaningful insights into model behavior [page::1][page::2]. A large-scale risk-attribution study on Form 10-K risk sections demonstrates practical interpretability for risk management and model diagnostics, highlighting how risk headings can be decomposed into component contributions with robust stability and fidelity [page::25][page::27].
