Development of a Corpus-Grounded AI Self-Learning System for Arabic Language Learning

Abstract

This study is kicked off from the fact that the level of Arabic language proficiency nowadays among students at various levels of education, especially in universities, is facing serious problems. They face more critical problems particularly in the four language skills; reading, speaking, listening and writing. Therefore, this study tries to develop a corpus-grounded artificial intelligence (AI) self-learning system for two compulsory Arabic courses at Universiti Sains Islam Malaysia (USIM) to assist solving this problem. Firstly, a corpus consisting of these Arabic courses text was developed. They were collected, cleaned, segmented into pedagogical items, organized with metadata, and represented semantically for retrieval. The resulting pedagogical corpus comprised 524 searchable items, including 434 from Arabic Language II (Communication) course and 90 from Arabic Language III (Social Science) course. To select the AI development design, 43 authentic grammar, morphology, and reading-comprehension questions were used to compare two locally deployable Arabic-specialised models with a general-purpose model using retrieval-augmented generation (RAG). The configurations achieved 37.2%, 32.6%, and 67.4% accuracy, respectively, supporting the selection of RAG design to be implemented. An AI Self-learning System was successfully developed and ready to use by specific students who take these two compulsory courses. The implemented system uses semantic retrieval to ground generative AI responses in course-specific corpus evidence and provides study assistance with source traceability, dynamic practice generation, and diagnostic support. The study demonstrates an implemented pathway from Arabic educational materials to a functioning corpus-grounded RAG system. The findings establish technical and functional implementation rather than learner learning gains.