Longitudinal development of L2 speaking accuracy in AI-enhanced smart class

Evidence from an error-annotated learner corpus

Bibliographic Data

ID21535676
AuthorsMengjia Liu (0009-0002-2641-2639, National University of Malaysia), Harwati Hashim (0000-0002-8817-427X, National University of Malaysia, corresponding author), Nur Ainil Sulaiman (0000-0001-6212-7494, National University of Malaysia)
Year2026
Volume140
Pages104048
Publication date2026-07-01
Peer ReviewedYes
Open AccessYes
TypeARTICLE
VenueSystem (JOURNAL)
Journal identifiersISSN: 0346-251X • E-ISSN: 1879-3282
PublisherElsevier BV (PUBLISHER)
DOI10.1016/j.system.2026.104048
OpenAlexW7160014736
LanguageEN
References cited33

This study investigates the longitudinal development of L2 speaking accuracy in Chinese EFL learners within AI-enhanced smart classroom environments over extended periods. The research examines four key dimensions of speaking accuracy—lexical accuracy, lexical diversity, grammatical accuracy, and grammatical complexity—to establish empirical connections between sustained AI-mediated instruction and measurable linguistic improvement. A corpus-based longitudinal design tracked 100 s-year English majors across two academic semesters at three strategic measurement points. The study employed Error-Annotated Spoken Learner Corpus (EASLC), comprising 18,750 utterances with comprehensive error annotation. Statistical analyses included repeated-measures ANOVA, concordance analysis, and error pattern documentation to track developmental trajectories and error remediation patterns across proficiency levels. Results demonstrated significant improvements across all accuracy dimensions (ηp 2 = .44-.49), with grammatical complexity showing the largest gains (31.3%). Proficiency-differentiated trajectories revealed lower-proficiency learners achieved greater improvements (d = 1.42-1.78) compared to advanced learners (d = .48-.72). Morphosyntactic errors decreased 72-76%, while pragmatic errors persisted despite repeated AI feedback. Self-correction rates increased from 12% to 34%, indicating enhanced metalinguistic awareness. Cluster analysis identified three developmental profiles correlating with digital literacy and AI tool engagement patterns. This methodological integration revealed three distinct developmental profiles correlating with digital literacy and AI tool engagement patterns, and uncovered differential AI feedback responsiveness across linguistic domains—with form-focused features showing stronger remediation than discourse-pragmatic elements. The research informs optimal integration strategies for AI-powered tools in language instruction and contributes theoretical insights to technology-mediated second language acquisition research

Computational linguistics · Corpus linguistics · Language acquisition · Language development · Longitudinal study · AI in Service Interactions · Second Language Acquisition and Learning · Text Readability and Simplification

Tools

Open DOIOpen Access