Epimopilot : An LLM-infused Epistemic Emotion Support System to Boost Self-regulated Asynchronous Online Video Learning

Paper Title

Epimopilot: An LLM-infused Epistemic Emotion Support System to Boost Self-regulated Asynchronous Online Video Learning

Publication Info

• Topic area: Enhancing self-regulated asynchronous video learning through emotional and cognitive support.
• Keywords: Epistemic emotions, asynchronous learning, self-regulated learning, large language models (LLMs), danmaku, emotion recognition, knowledge graph, AI tutor, co-presence, cognitive scaffolding.

Background and Problem

• Problem / challenge: Asynchronous video learning platforms lack emotional and cognitive support, leading to isolation, unresolved confusion, and diminished engagement. Current systems fail to integrate real-time emotional cues with proactive interventions.
• Significance: Addressing these limitations can enhance learning persistence, engagement, and comprehension, making asynchronous platforms more effective and supportive.
• Motivation and related work: Prior research has explored danmaku comments and LLMs in education but has not effectively combined emotional recognition with adaptive, real-time support. This paper builds on these gaps to create a system that integrates emotional and cognitive scaffolding.

Solution

• Proposed approach: Epimopilot, a system that combines collective epistemic emotion recognition from danmaku comments with LLM-powered interventions to provide emotional and cognitive support during asynchronous video learning.
• Novelty:
  1. Transforming chaotic danmaku streams into structured cognitive and emotional insights.
  2. Leveraging LLMs for proactive, context-aware interventions based on collective emotional states.
  3. Integrating a knowledge graph and emotion distribution map for strategic learning.
• Procedure and key techniques:
  • Backend processes video content and danmaku comments using LLMs and BERT for emotion annotation.
  • Frontend includes three components:
    1. Epimo map: Combines a knowledge graph with an emotion distribution map.
    2. Epimo companion: Provides ambient emotional resonance and structured discussion insights.
    3. Epimo copilot: Offers just-in-time LLM-powered tutoring based on detected learning difficulties.

Results

• Concrete findings:
  • Significant improvements in knowledge comprehension (e.g., +7.82 points for Video B, p < .001) and recall (e.g., +7.81 points for Video B, p < .001) compared to the baseline.
  • Enhanced emotional presence (e.g., +2.23 points for Video B, p < .001) and learning persistence (e.g., +1.92 points for Video B, p < .001).
  • Reduced cognitive load in dimensions like frustration and effort.
• Advantage over baselines:
  • Outperformed a baseline system (basic video player with danmaku) in engagement, comprehension, and emotional support metrics.
  • Provided timely, context-aware interventions that prevented confusion from escalating into frustration.
• Experiments / evaluation:
  • Controlled within-subjects study with 32 participants (age 18–26).
  • Participants used both the baseline and Epimopilot systems across two videos (biology and AI topics).
  • Metrics included quiz scores, NASA-TLX cognitive load, and subjective evaluations.
• Limitations and future work:
  • Information density and cognitive load management need refinement.
  • Emotional expressivity calibration and long-term effects on self-regulation require further study.
  • Future directions include adaptive personalization, longitudinal studies, and expanded domain applicability.

Summary

Epimopilot is an innovative system that enhances asynchronous video learning by integrating collective epistemic emotion recognition with LLM-powered interventions. It provides emotional resonance through a virtual companion and cognitive scaffolding via an AI tutor and knowledge graph. A controlled study demonstrated significant improvements in learning outcomes, engagement, and emotional support compared to a baseline system. While effective, challenges in information density and long-term impact on self-regulation highlight areas for future research. This work offers a blueprint for emotionally intelligent, adaptive learning environments.