Study finds translation-based fine-tuning feasible for low-resource languages

A study evaluates the feasibility of translating non-English data into English to fine-tune existing BERT models, offering a resource-efficient alternative to developing native-language models. The research compares this approach against native-language performance across six NLP tasks using datasets from Bulgarian, Chinese, Dutch, Italian, and Russian.

• Translation-based fine-tuning was comparable or superior in 53.3 percent of cases across all settings.
• Gains were most frequent in Question Answering, Part-of-Speech Tagging, and Natural Language Inference.
• Performance declines were common in Named Entity Recognition and Hate Speech Detection.
• The approach is most effective for tasks relying on syntactic patterns and languages typologically close to English, such as Dutch.

The results demonstrate that translation-based fine-tuning provides a scalable path for extending NLP to low-resource languages while advancing linguistic inclusivity.