Beyond Chatbots: Enhancing Luxembourgish Language Learning Through Multi-agent Systems and Large Language Model

Authors

Nouzri S., EL Fatimi M., Guerin T., Othmane M., Najjar A.

Reference

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 15395 LNAI, pp. 385-401, 2025

Description

The intersection of Artificial Intelligence (AI) and education is transforming learning and teaching, with Generative AI (GenAI) and large language models (LLMs) offering new possibilities. AI and LLMs personalize learning through adaptive study guides, instant feedback, automated grading, and content creation, making resources more accessible and tailored to individual needs. Notably, LLM-based chatbots, such as OpenAI’s ChatGPT, serve as virtual assistants, ideal for language practice. However, these chatbots often limit themselves to teaching vocabulary through role-playing conversations or providing instant feedback based on model-generated content, which may lead to exposure to inaccuracies. This overlooks the holistic nature of Language Learning (LL), which requires pedagogy, effective methods, reliable content, and a supportive teacher-student relationship. Therefore, relying on a single chatbot is inefficient for the entire learning process. In this paper, a Multi-Agent System (MAS) is proposed, where each agent specializes in a specific function, working together to provide personalized, adaptive learning support. This approach breaks down the complex learning process into manageable parts. It employs the Business Process Model and Notation (BPMN), translated into agent-based modeling and LLMs to create dynamic, tailored learning environments. By simulating interactions similar to human tutoring, this model ensures real-time adjustments to meet each student’s evolving needs. Our project aims to address these limitations by using LL books with robust pedagogical resources as primary references. We focus on teaching Luxembourgish, adding complexity to our challenges as it is a low-resource language, ensuring a holistic learning experience. Our approach employs complex LLM workflows as multi-agent collaborations for reading, conversing, listening, and mastering grammar, based on GPT-4o, enhanced by Retrieval-Augmented Generation (RAG) and voice recognition features.

Link

doi:10.1007/978-3-031-77367-9_29