1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT)

The StereACuLT workshop aims to broaden and deepen the study of stereotypes in language technologies by foregrounding cultural context. From static word embeddings to contemporary transformer-based models, stereotypes encoded in language representations have remained a persistent challenge. Although widely used benchmarks such as StereoSet and CrowS-Pairs have driven progress in measuring and mitigating such biases, they are predominantly English-centric, limiting their validity and applicability in multilingual and multicultural settings.

The rapid deployment of large, multilingual language models further amplifies this limitation. Culture-agnostic assumptions and heuristics can lead to systematic failures, including under- or over-moderation, miscalibrated safety mechanisms, and misaligned value priors when models are applied across cultural contexts. Addressing stereotypes therefore requires not only technical solutions, but also culturally grounded definitions, evaluation protocols, and mitigation strategies.

This workshop seeks to accelerate principled, transparent, and replicable approaches for defining, localizing, evaluating, and mitigating stereotypes across cultures. We emphasize methods that explicitly incorporate culturally valid stereotype operationalizations, and that account for linguistic, social, and geopolitical variation. We are particularly interested in work that contrasts countries or communities sharing a language, distinguishes between diaspora and local populations, or examines how alignment interventions such as RLHF, RLAIF, or policy tuning behave under cultural shift.