Evaluating multimodal explainable adaptive learning in university EFL education: Learner outcomes, teacher oversight, and implementation conditions.

Abstract

Artificial intelligence (AI) offers substantial potential for personalized English as a Foreign Language (EFL) instruction, yet concerns remain regarding the transparency of adaptive recommendations, the fragmentation of learner data, and the preservation of teacher oversight. This study empirically evaluates the Multimodal Explainable Adaptive Learning Framework (MEALF), a pedagogical-computational framework operationalized through a prototype system that integrates textual, oral, visual, and behavioral learner information with explainable adaptive recommendations. Grounded in sociocultural theory, situated cognition, and human-centered AI, MEALF was designed to provide differentiated learning support while keeping pedagogical decisions interpretable and subject to teacher judgment. A 16-week partially randomized quasi-experimental mixed-methods study was conducted with 100 intermediate-level EFL learners and 15 teachers across three Algerian universities. Learners in the MEALF-supported condition showed larger proficiency gains and higher post-intervention engagement and metacognitive reading-strategy awareness scores than learners in the comparison condition. Exploratory multimodal analyses further revealed heterogeneous learner patterns that provided additional context for interpreting differences in learning needs and adaptive support. Teacher and qualitative evidence indicated that explanations and override mechanisms helped educators inspect and critically evaluate system recommendations, although teacher-level findings should be interpreted cautiously given the small sample. The findings provide preliminary, context-specific evidence that combining multimodal learner information with interpretable adaptive support may contribute to language development, engagement, and metacognitive reading-strategy awareness while preserving meaningful human oversight. However, the multi-component nature of MEALF, the quasi-experimental design, the small teacher sample, and the single-country context limit causal interpretation and broader generalizability. The study contributes an integrated empirical evaluation of multimodal, explainable adaptive learning by linking measurable learner outcomes with interpretable human-AI interaction and contextual implementation evidence in university EFL education.