Discovering Self-Regulated Learning Patterns in Chatbot-Powered Education Environment

TL;DR

This paper investigates self-regulated learning (SRL) in GenAI-powered education by introducing Gen-SRL, an annotation schema that maps student–chatbot interactions to 16 microlevel SRL actions across four macrolevels. Using 212 real-world English writing interactions from the RECIPE4U dataset and process mining, the study reveals that learners predominantly engage in task-execution (control) actions, with limited planning and reflection, and that SRL regulation patterns are nonsequential, featuring phase skipping and frequent self-loops. The contributions include the Gen-SRL schema, application of process mining to SRL in chatbot-mediated contexts, and evidence-based guidance for designing adaptive, personalized scaffolds and Socratic-style prompts to enhance metacognitive engagement. The findings challenge classical SRL theories in GenAI contexts and offer practical implications for building educational chatbots that actively support self-regulation and autonomous, reflective learning. Overall, the work advances SRL research by presenting a dynamic, context-aware approach to analyze and foster self-regulation in AI-enhanced learning environments.

Abstract

The increasing adoption of generative AI (GenAI) tools such as chatbots in education presents new opportunities to support students' self-regulated learning (SRL), but also raises concerns about how learners actually engage in planning, executing, and reflection when learning with a chatbot. While SRL is typically conceptualized as a sequential process, little is known about how it unfolds during real-world student-chatbot interactions. To explore this, we proposed Gen-SRL, an annotation schema to categorize student prompts into 16 microlevel actions across 4 macrolevel phases. Using the proposed schema, we annotated 212 chatbot interactions from a real-world English writing task. We then performed frequency analysis and process mining (PM) techniques to discover SRL patterns in depth. Our results revealed that students' SRL behaviours were imbalanced, with over 82% of actions focused on task execution and limited engagement in planning and reflection. In addition, the process analysis showed nonsequential regulation patterns. Our findings suggest that classical SRL theories cannot fully capture the dynamic SRL patterns that emerge during chatbot interactions. Furthermore, we highlight the importance of designing adaptive and personalized scaffolds that respond to students' dynamic behaviours in chatbot-powered contexts. More importantly, this study offers a new perspective for advancing SRL research and suggests directions for developing chatbots that better support self-regulation.

Figures (7)

• Figure 1: A reproduction of the theoretical SRL model, based on Pintrich's four-phase model pintrich2004.
• Figure 2: Overview of the proposed methodology, consisting of schema formulation and empirical implementation stages.
• Figure 3: Annotated snippet of a student–chatbot conversation showing SRL microlevel actions based on the Gen-SRL annotation schema.
• Figure 4: Macrolevel SRL process model generated using Disco with activity and path details set to 100% and 60%, respectively. Rectangles represent SRL phases; green/red nodes indicate start/end events. Node shade and edge thickness reflect frequency and transition counts, respectively. All values are absolute frequencies. The full transition frequency matrix is provided in Appendix B for reference.
• Figure 5: Microlevel SRL process model generated using Disco with activity and path details set to 100% and 17.8%, respectively. Visual elements follow Figure \ref{['fig:macro_recipe4u']}; all values represent absolute frequencies. The full transition frequency matrix is provided in Appendix B for reference.