Toward Culturally Grounded Natural Language Processing

A new research paper by Sina Bagheri Nezhad, titled 'Toward Culturally Grounded Natural Language Processing,' synthesizes findings from over 50 studies published between 2020 and 2026. The analysis reveals a critical gap in current AI: strong multilingual capability does not equate to cultural competence. The paper examines issues like performance inequality across languages, flawed cross-lingual transfer, and critiques of benchmark design, showing that models can excel on tests like Global-MMLU or CDEval while still flattening local norms and misinterpreting culturally grounded cues.

Key technical factors identified include tokenization, the language of prompts, and the use of translated benchmarks, which all materially affect outcomes. The paper reviews newer, culturally-aware benchmarks like WorldValuesBench, CulturalVQA, and DRISHTIKON, which highlight these shortcomings. Nezhad argues the field must shift from viewing languages as spreadsheet rows to modeling 'communicative ecologies'—the full context of institutions, scripts, and communities where language is used.

Based on this synthesis, the paper proposes a concrete research agenda for culturally grounded NLP. This agenda centers on five pillars: creating richer contextual metadata for training data, implementing culturally stratified evaluation, adopting participatory alignment with local communities, modeling within-language variation, and pursuing multimodal, community-aware design. The goal is to build AI systems that understand not just the words, but the cultural and social fabric in which they are embedded.