LANGUAGE AND SOCIAL BIAS CATEGORIES ASSOCIATED WITH SENTENCE LENGTH ASYMMETRY IN MULTILINGUAL STEREOTYPE EVALUATION BENCHMARKS

This analysis examines whether corresponding multilingual stereotype benchmarks provide comparable textual conditions for international language technology evaluation. Using 3,353 complete English and French sentence pairs from a benchmark extension released in 2022, descriptive comparisons assess normalized character length asymmetry across languages, social categories and stereotype orientations. Among 1,676 matched identifiers, mean asymmetry is 0.03737 in English and 0.02869 in French, yielding a mean paired difference of negative 0.00868. However, French asymmetry exceeds English asymmetry in 43.74 percent of matched pairs, revealing heterogeneity beneath the aggregate ordering. Restricting comparisons to 1,535 identifiers with identical category annotations preserves the lower French mean. Category rankings differ across languages, while aggregate orientation comparisons reverse direction. These patterns distinguish correspondence between records from agreement in classification and similarity in sentence construction. The contribution refines benchmark comparability as a measurable, multidimensional consideration rather than an assumed consequence of translation. For multilingual evaluation governance, the findings support retaining correspondence identifiers, documenting annotation differences and reporting matched distributions alongside language specific summaries.