Using Gaze Transition Entropy to Detect Classroom Discourse in a Virtual Reality Classroom

The authors of this paper include Philipp Stark, affiliated with the Human-Computer Interaction program at the University of Tübingen, Germany; Alexander Jung, also from the University of Tübingen; Jens-Uwe Hahn, who holds a Ph.D. from the University of Tübingen and is currently a professor at the Stuttgart Media University, focusing on rendering and visualization; Enkelejda Kasneci, who received a Ph.D. in Computer Science from the University of Tübingen and is now a professor of Human-Centered Technology at the Technical University of Munich; and Göllner Richard, whose research involves virtual reality, teaching quality, and more.

Abstract

• Explore the use of gaze transfer entropy to detect classroom discourse events in virtual reality (VR) classrooms.
• Using data from laboratory experiments, distinguished between teacher-centered classroom discourse events (questions, hand raises, answers) and teacher explanations.
• Employed multilevel regression models, where two entropy metrics effectively differentiated between these events and different levels of classroom participation indicated by virtual student hand raises.
• The potential of gaze entropy was demonstrated in logistic regression models, achieving 67% accuracy in predicting the two events.
• By analyzing transfer and static entropy, the study aimed to reveal different gaze patterns associated with learning events in virtual classrooms.
• Results contribute to the research and development of VR scenarios that simulate effective learning environments.

1. Introduction

• The classroom as the primary environment for student learning.
• VR classrooms can provide immersive learning experiences, simulating interactive classroom discourse.
• Classroom discourse refers to the collaborative learning process among students, characterized by active participation and behavioral engagement.
• This study aims to detect visual attention of students during multiple discourse events during the course.
• Utilizing gaze transfer entropy to investigate visual attention in VR classrooms.
• Research questions:
  • R1: Can transfer and static entropy be used to differentiate between animated classroom discourse events (teacher questions, hand raises, student answers) and teacher explanations (teaching course content) in VR classrooms?
  • R2: Does the predictive value of the two entropy metrics (transfer entropy and static entropy) depend on different levels of student participation indicated by hand raises?

2. Related Research

• Student attention behaviors in real classrooms.
• Research on VR classrooms: teacher roles, student attention, social information, learning, virtual environment design, student seating posture, class size.
• Using eye-tracking to study visual attention.
• Calculation methods for gaze transfer entropy.
• Applications of gaze transfer entropy in educational research.

3. Method

• Experiment and Sample: VR laboratory experiment with sixth-grade students from Baden-Württemberg, Germany.
• Data Aggregation and Metrics: Defined AOIs, created duration and transition data, computed transfer and static entropy.
• Data Analysis: Multilevel linear regression models and logistic regression models.

4. Results

• Differences in transfer entropy and static entropy between events.
• Transfer entropy had higher predictive value under 80% and 20% hand raise conditions.
• Static entropy had higher predictive value only under 80% hand raise condition.
• Logistic regression models using both entropy metrics predicted events with 67% accuracy.

5. Discussion

• Gaze transfer entropy can effectively detect classroom discourse events in VR classrooms.
• The level of virtual student participation affects the predictive value of gaze transfer entropy.
• Limitations of the study: Standardized settings, technical challenges.
• Future research directions: More complex VR classroom scenarios, mobile eye-tracking, integrating other metrics.

6. Conclusion

• Gaze transfer entropy is a valuable indicator for detecting classroom discourse events in VR classrooms.
• Transfer and static entropy can differentiate between discourse events and teacher explanations.
• The study results emphasize the importance of designing virtual avatars in educational environments.

Q: What specific research methods were used in the paper?

• Experimental Design: The experiment employed a virtual reality (VR) classroom environment to simulate scenarios of teachers delivering lectures and students participating in class discussions. The participants were 240 middle school students, randomly assigned to different levels of engagement (percentage of virtual students raising hands) in experimental groups.
• Data Collection: Eye-tracking data were collected using the HTC-VIVE Pro Eye head-mounted display and Tobii eye tracker.
• Data Analysis:
  • Eye Movement Data Processing: Gaze-ray casting technology was used to obtain the intersection points between the participant's gaze point and virtual objects, defining the virtual teacher, lectern, and virtual students as areas of interest (AOI).
  • Entropy Calculation: The fixation duration on each AOI and transitions between AOIs were calculated, and the transition entropy and stationary entropy for each 30-second interval were computed.
  • Statistical Analysis: Hierarchical linear regression models were used to analyze the relationship between entropy values and classroom events (teacher's lecture vs. class discussion) and student engagement levels. Logistic regression models were used to predict classroom events and assess the model's prediction accuracy.

Q: What are the main research findings and outcomes?

• Entropy Differences: There were significant differences in transition entropy and stationary entropy between teacher's lecture and class discussion events. The entropy values were higher during class discussions, indicating more visual exploration and a more uniform distribution of visual attention among participants in these events.
• Impact of Engagement Level: The percentage of virtual students raising hands affected the predictive value of transition entropy and stationary entropy. Under 80% and 20% hand-raising conditions, the predictive value of transition entropy was higher, indicating more visual exploration behavior. Stationary entropy had higher predictive value only under the 80% hand-raising condition, suggesting a more uniform distribution of visual attention under those conditions.
• Prediction Model: A logistic regression model using transition entropy and stationary entropy as input variables could predict classroom events with 67% accuracy, indicating that entropy values have significant predictive power in distinguishing between teacher's lecture and class discussion events.

Q: What are the current limitations of this study?

• Experimental Environment: The study was conducted in a highly standardized VR environment, lacking the complexity and interactivity of real classrooms.
• Data Collection: The accuracy and precision of VR eye trackers may affect the accuracy of AOI duration and transitions.
• AOI Definition: Defining AOIs based on virtual objects may lead to missed detections of other gaze transitions in the environment.
• Time Intervals: Due to data loss, the study used a longer 30-second interval, which may not capture immediate gaze behaviors during events.
• Model Prediction: When using 3 AOIs to calculate entropy, the model's prediction accuracy decreased, indicating the need for more AOIs to capture a more comprehensive gaze behavior.