.png){kind=icon}
.png){kind=icon}
Return
From declarative knowledge to procedural fluency: a generative AI tutor for Arabic agreement rules
.png){kind=icon}
PRE
AI
DOI：10.1080/09588221.2026.2685745.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Despite extensive development of computer-assisted language learning (CALL) tools and Intelligent Tutoring Systems (ITS) addressing grammar, learners often still struggle to apply complex morpho-syntactic rules in spontaneous speech. In particular, Arabic poses challenges in agreement (gender, number, case) that traditional instruction and rule-based feedback have only partially overcome. Moreover, recent reviews reveal that generative-AI research in language learning has focused overwhelmingly on written English tasks, leaving speaking and less-studied languages largely underexplored. To fill this gap, we conducted a longitudinal mixed-methods quasi-experiment comparing a GPT-4–based dialogic tutor to conventional explicit instruction for non-native Arabic learners’ oral mastery of agreement rules. Participants (N = 60 intermediate AFL learners) were assigned by intact class to either an AI-mediated speaking practice environment or traditional drill-based instruction. Pre-, post-, and delayed-post oral tests (elicited speech tasks targeting verb–subject, adjective–noun, and subject–predicate agreement) were analysed for accuracy in obligatory contexts, error density per 100 words, and fluency metrics (speech rate, pause ratio, response latency). System log data (feedback events, response times) and learner questionnaires (anxiety, perceived usefulness) provided additional insights. Results from ANCOVAs and mixed-effects models suggested that the AI group outperformed the control on agreement accuracy and showed greater improvements in fluency indices, with these gains largely maintained at a 4-week delay. Error density declined more sharply in the AI condition. Learning analytics indicated that log-derived features (e.g. accuracy, focus, and time-on-task) significantly predicted individual gains. Learners interacting with the AI tutor reported lower speaking anxiety and high technology acceptance, consistent with broader evidence that AI-mediated speaking support can enhance enjoyment and willingness to communicateLearners interacting with the AI tutor reported lower speaking anxiety and high technology acceptance, echoing Zhang et al. (2024) findings of boosted enjoyment and willingness to communicate under an AI speaking assistant. Qualitative comments further suggested that the dialogic agent may have functioned as a safe, personalised practice space. In sum, this study suggests that a generative-AI dialogic tutor may support the transition from explicit rule knowledge to more fluent use of Arabic agreement, thereby potentially scaffolding the proceduralisation of grammar. This work helps address the dearth of speaking-focused GenAI research, is consistent with cognitive load theory, and may extend skill-acquisition accounts by suggesting a possible role for GenAI in lowering the affective filter and enhancing noticing. Pedagogically, we frame our system as an Adaptive Learning Ecosystem that dynamically modulates task difficulty and feedback. We conclude with recommendations for curriculum designers on ethically integrating AI tutors – from scaffolding prompts to protecting student data – to harness GenAI affordances while preserving language-specific complexity.
Keywords:
Generative AI in CALL
Arabic agreement
oral morpho-syntactic accuracy
proceduralisation of grammar
speaking anxiety
Journal
IF:
6.6
Papers:
3.0K
Citations:
5.2K