Although a growing body of research reports positive effects of AI chatbots on learning support, some studies caution that learners may rely on chatbot answers in a fragmented manner, potentially weakening conceptual understanding and critical thinkin...
Although a growing body of research reports positive effects of AI chatbots on learning support, some studies caution that learners may rely on chatbot answers in a fragmented manner, potentially weakening conceptual understanding and critical thinking. Against this backdrop, the present study investigated whether students can actively construct knowledge in a science-education chatbot environment. To this end, we analyzed students’ question types and question–answer interactions in a science chatbot and compared the resulting patterns with findings and interpretations from prior studies.
The dataset consisted of log-based conversation records collected over approximately one year from a single high school, comprising 7,907 question–answer pairs across 2,419 user sessions. Among these, single-turn exchanges accounted for 66.4%, whereas multi-turn sequences occurred in 33.5% of sessions: 2-turn (16.5%), 3-turn (7.3%), 4-turn (4.1%), 5-turn (2.1%), and 6+ turns (3.6%). For question-type analysis, we adopted the coding framework proposed by Kim (1998). For interaction analysis, we applied the Interaction Analysis Model (IAM) developed by Gunawardena et al. (1997). To establish reliability, two researchers independently coded a subset of the data, achieving Cohen’s kappa values of 0.70 and 0.64, after which the remaining data were coded by the primary researcher.
Across all 7,907 questions, the distribution of question types was as follows: Recall (31.5%), Reframe (8.4%), Apply (4.7%), Extend (8.8%), Contradiction (5.6%), and Irrelevant (38.0%). In addition, a novel category emerged in which students requested the generation of supplementary learning resources (3.0%), which was coded as Request. Compared with prior research analyzing student questions in classroom settings (Recall: 23.2%; Contradiction: 3.4%), the proportions of Recall and Contradiction were higher in the chatbot environment, suggesting that students may find it easier to externalize knowledge gaps and inconsistencies as questions outside the classroom context.
For the first question in single-turn exchanges and in multi-turn sequences, the distribution was Recall (65.8%), Reframe (13.0%), Apply (6.9%), Extend (4.2%), Contradiction (5.2%), and Request (4.9%). Beyond the second turn, question types shifted systematically by turn. The proportion of Recall gradually decreased to 21.5% in 6+ turns, whereas Contradiction increased to 23.8% in 6+ turns. Extend rose sharply at the second turn and remained high (24.5%–32.6%) thereafter. These patterns suggest that students did not merely accept chatbot answers dependently; rather, they increasingly expressed inconsistencies and pursued elaboration, indicating attempts at active knowledge construction during multi-turn interaction.
For IAM-based interaction analysis, we examined 2,813 analyzable question–answer exchanges drawn from 490 sessions with five or more turns (excluding unanalyzable questions). The distribution across IAM phases was Phase 1 (48.9%), Phase 2 (26.7%), Phase 3 (20.1%), Phase 4 (3.2%), and Phase 5 (1.2%). Compared with prior studies of human–human online interaction (Phase 2: 2.4%–20.6%; Phase 3: 1.9%–12.7%; Phase 4: 0%–1.0%), Phases 2–4 were higher in the chatbot context, implying more frequent attempts at social knowledge construction within question–answer interaction. However, Phase 5 showed no meaningful difference from prior studies (0%–1.9%), suggesting that reaching a level of negotiated agreement and application remains difficult even in chatbot-mediated interaction.
Overall, the findings indicate that the chatbot environment may facilitate students’ question expression and holds potential as an interactive tool that supports active knowledge construction. Nevertheless, the study is limited by the restricted scope of the sample and by the use of an analytic framework not specifically optimized for chatbot contexts, which constrains interpretive validity. Future research should develop and validate question analysis frameworks tailored to chatbot-based learning environments.