Decoding Student–Chatbot Dialogues: How Interaction Structure Is Associated with Learning Gains in AI-Assisted Programming

The study examines how secondary school students interacted with an AI-powered educational chatbot, MyBotBuddy, while working on a programming task, and how observed dialogue structures were associated with differences in pre- to post-test performance. Fifty students first completed an unassisted pre-test, then attempted a chatbot-supported programming task, and finally completed an unassisted post-test. Based on score change, students were grouped into learning gain, no gain, and learning loss categories. Dialogue transcripts were analyzed using Epistemic Network Analysis to identify co-occurring discourse patterns, alongside descriptive sentiment analysis to characterize lexical tone. Students in the learning gain group showed more connected multi-turn patterns involving solution attempts, feedback uptake, knowledge-related contributions, and clarification following feedback. In contrast, the no gain and learning loss groups showed less iterative and less systematically connected interaction structures. Average sentiment polarity differed only slightly across groups and is interpreted cautiously because the dialogue was technical and programming focused. The findings are associational and exploratory rather than causal and suggest that learner engagement with a chatbot may be more informative than interaction frequency alone. We discuss implications for educational chatbot design, especially the potential value of multi-turn scaffolding and reflective prompting, while outlining the need for future validation, baseline-controlled analyses, and experimental work.