Multilingual Conversational Tutoring Systems Using Transformer-Based Contextual Understanding in STEM Learning

Abstract

The lack of scalable, language-inclusive tutoring systems presents a significant challenge in delivering equitable STEM education to multilingual learners. Traditional intelligent tutoring systems are typically monolingual and struggle to provide adaptive, conversational support that accommodates linguistic diversity and domain complexity. This study aims to develop a multilingual conversational tutoring system that leverages transformer-based contextual understanding to deliver personalized STEM instruction across multiple languages. The proposed framework integrates multilingual transformer models (mT5 and mBERT), a context-aware dialogue manager, and a domain-specific reasoning module capable of symbolic computation for solving math and science problems. A custom multilingual STEM dialogue corpus was constructed using educational forums, open resources, and synthetically generated dialogues in English, Spanish, Hindi, and Mandarin. The system was fine-tuned using weighted language sampling and further optimized via reinforcement learning from human feedback. Experimental results demonstrate robust cross-lingual performance: BLEU-4 scores of 0.762 (English), 0.733 (Spanish), 0.712 (Hindi), and 0.694 (Mandarin); contextual relevance exceeding 85% across languages; and STEM Conceptual Accuracy reaching up to 93.1%. Human evaluations yielded an average User Satisfaction Score of 4.5/5 in English and above 4.2 in all languages. This work contributes a scalable, real-time tutoring architecture that bridges language gaps in STEM education. The system shows strong potential for deployment in multilingual classrooms and remote learning environments, offering personalized, high-quality support to learners worldwide