Speakers
Description
This study examines learner interaction data from an AI-driven extensive reading system that generates stories levelled to reader proficiency. Behavioural measures such as reading speed, completion, and rereading are analysed alongside learner comments and ratings to identify patterns associated with successful and problematic texts. Findings provide insight into how well adaptive generation aligns with learner needs and where mismatches occur. Attendees will have the opportunity to try the system and generate an AI-driven story firsthand.
Short summary
How well can AI working with students generate level-appropriate stories for Extensive Reading? This study examines interaction data from over 600 AI-generated stories, combining behavioural measures (reading speed, completion, rereading) with learner comments and ratings. Clustering and sentiment analysis reveal patterns associated with successful texts and identify where adaptive levelling performs well, and where it struggles. Attendees will have the opportunity to try the system and generate an AI-driven story firsthand.
Abstract
This study examines learner interaction data from an AI-driven extensive reading system that generates stories adaptively levelled to reader proficiency. The system adjusts linguistic complexity at the point of generation rather than selecting from a fixed corpus, enabling investigation of levelling accuracy across the proficiency spectrum. Recent work has highlighted both the potential of AI to support extensive reading and the need to evaluate text suitability and levelling accuracy (Authors, 2025).
Over 600 stories have been generated to date, with some read repeatedly by learners. The system collects fine-grained, page-level interaction data (including time/page) alongside learner comments and ratings. This enables examination of behavioural measures (reading speed, consistency, and abandonment) together with affective responses.
In extensive reading, sustained engagement with accessible texts is central to learning outcomes (Day & Bamford, 1998, 2002). Reading speed profiles offer one window into this: elevated speed may indicate superficial interaction, while reduced speed may reflect increased cognitive load. Completion at a stable pace suggests a text is both accessible and engaging.
Clustering analysis groups reading sessions according to behavioural patterns such as completion rate, speed consistency, and rereading frequency. Sentiment analysis of learner comments identifies patterns associated with successful or problematic texts. These findings are examined against intended text levels to evaluate levelling accuracy, with particular attention to performance at the lower and upper ends of the proficiency range where adaptive systems typically struggle most.
Keywords
LLMs
Extensive Reading
AI
Learner-Generated Content
References
Author1, & Author2 (2025). Extensive reading and AI: This could be great. In L. Ohashi, M. Hillis, & R. Dykes (Eds.), Artificial intelligence in our language learning classrooms (pp. 58-77). Candlin & Mynard. https://doi.org/10.47908/38/3
Day, R. R., & Bamford, J. (1998). Extensive reading in the second language classroom. Cambridge University Press.
Day, R. R., & Bamford, J. (2002). Top ten principles for teaching extensive reading. Reading in a Foreign Language, 14(2), 136–141.
| Scheduling preference | Anytime on Saturday |
|---|---|
| Title | AI+Student Generated Extensive Reading: What Learner Behaviour Reveals |