Who Let the Dogs Out? Auditing Dogwhistle Coverage in Hate Speech Benchmarks Across Target Groups

Audited dogwhistle coverage across two widely-used hate speech benchmarks (HateXplain, Measuring Hate Speech), mapping an existing 340-entry dogwhistle glossary onto a four-level coding sophistication taxonomy across 59.6k posts. Identified annotation failure (mislabeling dogwhistle content as non-hateful) and collection failure (entire dogwhistle categories missing from the data) as distinct, unevenly distributed failure modes — e.g., race and LGBTQ+ subgroups showing the largest gaps — and released the mapped glossary and pipeline publicly.

Recommended citation: (2026). "Who Let the Dogs Out? Auditing Dogwhistle Coverage in Hate Speech Benchmarks Across Target Groups." EMNLP WOAH 2026.