SemEval-2026 Task 6: CLARITY – Unmasking Political Question Evasions

AbstractThis paper presents CLARITY, the SemEval-2026 shared task on detecting and classifying evasive responses in political discourse. The task is grounded in an expert-designed two-level taxonomy and a benchmark dataset of question-answer pairs from U.S. presidential interviews, requiring systems to distinguish clear from evasive responses at a coarse level and identify one of nine fine-grained evasion strategies at a fine-grained level. With 124 registered teams and over 1,400 combined valid submissions, the task attracted broad participation spanning a wide range of methodological approaches, from fine-tuned encoder models to multi-stage large language model pipelines. Analysis of submitted systems reveals that hierarchical exploitation of the taxonomy and chain-of-thought prompted LLMs were the most effective strategies, while fine-grained evasion classification remained a substantially harder and largely unsolved challenge. CLARITY advances the study of strategic ambiguity in political language as a formal NLP benchmark and highlights key open problems in computational discourse analysis.