AFRILANGTUTOR: Advancing Language Tutoring and Culture Education in Low-Resource Languages with Large Language Models
Researchers have developed AFRILANGTUTOR, a novel approach to language learning for low-resource African languages. This system utilizes a new dataset, AFRILANGDICT, comprising nearly 200,000 African language-English dictionary entries, to generate extensive question-answer pairs for training AI tutors. The resulting AFRILANGEDU dataset, with over 78,000 multi-turn examples, was used to fine-tune Llama-3-8B-IT and Gemma-3-12B-IT models across ten African languages. Evaluations demonstrated that these fine-tuned models significantly outperform their base versions, with combined Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) yielding the most substantial improvements. AI
IMPACT Enables AI-powered language education for underserved linguistic communities, potentially preserving cultural heritage and improving access to information.