Introduction
The pedagogical strategy of "learning by teaching"—where a student reinforces their own knowledge by explaining concepts to others—has long been recognized as a powerful tool for deep learning. However, implementing this strategy through virtual agents often results in rigid, highly structured interactions, such as filling out concept maps. These constraints limit natural expression and fail to capture the complexities of subjects like history or literature. StuBot addresses this by leveraging Machine Reading Comprehension (MRC) to allow for free-form, unstructured student input, providing adaptive feedback that encourages deeper reflection.
System Architecture
StuBot is designed as a "virtual student" that learns from the user. Its architecture comprises three primary components:
- Answering Module: Utilizing KorBERT, a Korean-optimized BERT model, StuBot performs extractive MRC. It treats the user's free-form explanation as the context and attempts to answer ten predefined factoid questions.
- Self-grading Module: To assess its own "learning," StuBot compares its generated answers against a ground-truth key. It employs a Support Vector Machine (SVM) classifier that evaluates Jaccard similarity and the MRC model's confidence score to determine correctness.
- Messaging Module: When the agent fails to answer a question correctly, it identifies the missing information and prompts the user for further clarification using dialogue templates.
The Iterative Learning Process
The core of the StuBot experience is an iterative loop consisting of preparation, presentation, and feedback. Unlike traditional UX design that prioritizes frictionless interaction, StuBot intentionally introduces "productive friction." By asking for more detail on specific sub-topics, the agent forces the user to revisit the learning material and refine their mental models.
If the machine cannot find an answer within the student's text, it serves as a signal that the explanation was insufficient. This framing of the agent as a student preparing for an exam triggers a sense of responsibility in the user, motivating them to provide clearer and more comprehensive explanations.
Evaluation and Results
A within-subject study involving 20 participants compared StuBot’s adaptive feedback against a baseline without feedback. The results demonstrated significant improvements in learning outcomes:
- Learning Performance: Participants using StuBot scored significantly higher on post-session quizzes, particularly in descriptive ($p < .001$) and comprehensive ($p = .022$) questions.
- Immersion: StuBot achieved higher immersion scores, with users reporting a strong sense of "transportation"—feeling as though they were genuinely interacting with a student.
- Robustness: Interestingly, the system's occasional errors (a ~9% error rate in the MRC model) did not negatively impact the user experience. Users interpreted the bot's failure to understand as a sign of the "student" being slow, which further encouraged them to rephrase and clarify their explanations.
Discussion and Future Work
StuBot demonstrates that MRC can serve as an effective proxy for understanding in "learning by teaching" scenarios. The "inconvenient interaction" paradox suggests that challenging the user can lead to better pedagogical outcomes than a perfectly smooth interface.
Future research will explore transitioning from extractive to generative models (such as GPT-based architectures) to better handle semantic matching and synonyms. Additionally, expanding the system's application to diverse domains like science and mathematics, and optimizing the cognitive load to prevent user fatigue, remain key areas for development.
Summary updated