DeltaSHAP: a Shapley Value Framework for Interpreting Political Ambiguity

AbstractPolitical ambiguity and response clarity have become increasingly important research topics in computational social science and natural language processing. In this paper, we present a solution to the SemEval 2026 Task 6 “Clarity” Challenge. We propose a novel framework that employs TF–IDF representations and Shapley-value–based feature selection for multi-class classification. Shapley-based feature importances are used both for post-hoc explanation and as an active mechanism for label-specific vocabulary selection. For each label, features exceeding a predefined threshold are retained, label-specific vocabularies are filtered through set differences, and independent one-versus-all classifiers are trained using specific features. Experimental results show that threshold tuning substantially impacts performance, with the best performance achieved at intermediate threshold values. Our findings demonstrate that using the game-theoretic feature selection provides an interpretable approach to clarity classification, offering a flexible methodology for ambiguity-sensitive text analysis.