Who Brought Easter Eggs to Eid? Cultural Translation of Math Word Problems Across Seven Languages
Suchdev, Parisa
Suchdev, Parisa
Citations
Altmetric:
License
License
Abstract
Large language models (LLMs) are increasingly used to generate educational materials, including math word problems (MWPs) tailored to students� linguistic and cultural contexts. When narratives reflect familiar names, foods, currencies, and everyday settings, students may focus more on mathematical reasoning and less on interpreting unfamiliar contexts. Yet cultural adaptation by LLMs also introduces risks: changes may be superficial, stereotypical, or inaccurate, increasing rather than reducing cognitive load. This study examines how LLMs operationalize cultural knowledge when translating culturally embedded entities in MWPs. We prompt three frontier models, Claude Opus 4, GPT-4.1, and Gemini 2.5 Pro, to culturally translate 60 English MWPs from a culturally filtered subset of GSM-8K into seven target languages: Hindi, Bengali, Punjabi, Urdu, Sindhi, Italian, and Sicilian. Each translation is framed through the persona of an elementary school math teacher in a specified country. We extract culturally salient entities from the English problems, align them with translated counterparts, and classify each transformation as preserved, localized, type changed, generalized, or missing, producing a multilingual dataset of 6,489 entity-level transformations. Across five research questions, we find that entity types lie on a preservation-to-localization spectrum: life events, vehicle parts, and plant names are often preserved, while person names, currencies, and newspaper titles are almost always localized. Adaptation patterns cluster by national context, with the closest language pairs corresponding to shared country settings. Models frequently agree on the action applied to an entity but not on the specific replacement value. They also exhibit distinct adaptation personalities: Claude is most interventionist, GPT most conservative, and Gemini most likely to alter entity categories. Finally, all language-model combinations show value entropy collapse, indicating reduced diversity in translated cultural values relative to the English source. These findings show that model choice is itself a cultural choice and underscore the need to audit LLM-based educational adaptation for narrowing, misattribution, and culturally unstable outputs.
Description
Date
1/1/2026
Student Status
Graduate Student
Journal Title
Journal ISSN
Volume Title
Type of presentation
Oral Presentation
Collections
Research Projects
Organizational Units
Journal Issue
Citation
DOI
Advisor(s)
Department
Program/Major
Computer Science
College/School
College of Engineering and Mathematical Sciences
Organization
Research Category
Social Science (Religion)