---
id: wiki-2026-0508-emotionally-intelligent-tutoring
title: Emotionally Intelligent Tutoring Systems (EITS)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [EITS, affective tutor, AutoTutor, emotionally aware ITS]
duplicate_of: none
source_trust_level: A
confidence_score: 0.93
verification_status: applied
tags: [edtech, ai-tutor, affective-computing, eits, emotion, learning, pedagogy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: Tutoring system / LLM
---

# Emotionally Intelligent Tutoring Systems (EITS)

## 매 한 줄
> **"매 student 의 cognitive + emotional 의 동시 의 sense + respond 의 ITS"**. 매 frustration / boredom / confusion / engagement 의 detect → 매 strategy adjust. 매 famous: AutoTutor (Graesser), Affective AutoTutor. 매 modern: 매 LLM tutor + 매 facial / voice + 매 adaptive prompt.

## 매 핵심

### 매 emotion of learning (D'Mello)
- **Engagement**: 매 best.
- **Confusion**: 매 productive (zone of proximal).
- **Frustration**: 매 productive 의 X — 매 detect.
- **Boredom**: 매 challenge ↑.

### 매 affect detection
- **Behavioral**: 매 click, dwell, error.
- **Facial**: 매 brow furrow.
- **Posture**: 매 lean.
- **Speech**: 매 hesitation.
- **Self-report**: 매 emoji slider.

### 매 response strategy
- **Frustration** → 매 hint, scaffold.
- **Boredom** → 매 challenge ↑, novelty.
- **Confusion** → 매 dwell, allow.
- **Engagement** → 매 maintain.

### 매 응용
1. **Math tutor**: 매 step-by-step.
2. **Language**: 매 conversation practice.
3. **Programming**: 매 debug help.
4. **Adaptive learning**: 매 LMS.
5. **Reading**: 매 comprehension.

## 💻 패턴

### Affect detection (multimodal)
```python
class AffectDetector:
    def __init__(self):
        self.face = FacialAnalyzer()
        self.behavior = BehaviorTracker()
    
    def detect(self, frame, log):
        return {
            'engagement': 0.6 * self.face.attention(frame) + 0.4 * self.behavior.click_rate(log),
            'frustration': self.face.brow_furrow(frame) + self.behavior.delete_count(log),
            'confusion': self.behavior.idle_time(log),
            'boredom': self.face.yawn_count(frame) + self.behavior.skip_count(log),
        }
```

### Strategy selector
```python
def select_strategy(affect, mastery):
    if affect['frustration'] > 0.7:
        return 'hint_with_encouragement'
    if affect['boredom'] > 0.6 and mastery > 0.7:
        return 'increase_difficulty'
    if affect['confusion'] > 0.5 and mastery < 0.5:
        return 'review_prerequisite'
    if affect['engagement'] > 0.7:
        return 'maintain_flow'
    return 'check_in'
```

### LLM tutor with affect prompt
```python
def llm_tutor_response(student_msg, affect, history):
    affect_str = f"Frustration: {affect['frustration']:.1f}, Engagement: {affect['engagement']:.1f}"
    
    prompt = f"""You are a patient, emotionally-aware tutor.
Student affect: {affect_str}
{'IMPORTANT: Student frustrated — validate first, then small step.' if affect['frustration'] > 0.6 else ''}
{'IMPORTANT: Student bored — pivot to challenge.' if affect['boredom'] > 0.5 else ''}

History: {history[-3:]}
Student: {student_msg}

Response:"""
    return llm.generate(prompt)
```

### Productive vs unproductive confusion (D'Mello)
```python
def classify_confusion(affect_history, performance):
    duration = affect_history.confusion_duration()
    progress = performance.recent_correct_rate()
    if duration > 60 and progress < 0.3:
        return 'unproductive'  # 매 intervene
    return 'productive'  # 매 let dwell
```

### Mastery-aware (Bayesian Knowledge Tracing)
```python
class BKT:
    def __init__(self, p_init=0.1, p_learn=0.3, p_slip=0.1, p_guess=0.2):
        self.p_known = p_init
        self.p_learn, self.p_slip, self.p_guess = p_learn, p_slip, p_guess
    
    def update(self, correct):
        if correct:
            num = self.p_known * (1 - self.p_slip)
            denom = num + (1 - self.p_known) * self.p_guess
        else:
            num = self.p_known * self.p_slip
            denom = num + (1 - self.p_known) * (1 - self.p_guess)
        self.p_known = num / denom + (1 - num / denom) * self.p_learn
```

### Engagement intervention (recovery)
```python
def recover_engagement(disengaged_for_seconds):
    if disengaged_for_seconds < 30:
        return None
    if 30 <= disengaged_for_seconds < 120:
        return {'type': 'gentle_check_in', 'msg': 'Still with me?'}
    return {'type': 'pivot', 'msg': "Let's try something different."}
```

### Self-report (emoji slider)
```html
<div class="affect-checkin">
  How are you feeling?
  <button data-emotion="frustrated">😤</button>
  <button data-emotion="confused">🤔</button>
  <button data-emotion="bored">😴</button>
  <button data-emotion="engaged">🤩</button>
</div>
```

### Empathy response template
```python
EMPATHIC_OPENERS = {
    'frustrated': ["This one is tricky — that's a normal feeling.", "Let's slow down a bit."],
    'confused': ["I see what's confusing — let me explain differently.", "Good question."],
    'bored': ["Let me show you why this matters.", "Here's a more interesting twist."],
}

def open_response(detected_emotion):
    return random.choice(EMPATHIC_OPENERS.get(detected_emotion, ['']))
```

### A/B test (affect-aware vs not)
```python
def evaluate_eits(group_a_baseline, group_b_eits):
    return {
        'completion_a': mean(s.completed for s in group_a_baseline),
        'completion_b': mean(s.completed for s in group_b_eits),
        'satisfaction_a': mean(s.rating for s in group_a_baseline),
        'satisfaction_b': mean(s.rating for s in group_b_eits),
        'mastery_a': mean(s.mastery for s in group_a_baseline),
        'mastery_b': mean(s.mastery for s in group_b_eits),
    }
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Webcam OK | Multimodal facial + behavior |
| No camera | Behavioral + self-report |
| K-12 | Strong empathy emphasis |
| Higher-ed | Productive confusion tolerated |
| Adult learning | Less interruption |
| Mental health risk | Clinician escalation path |

**기본값**: 매 multimodal affect + 매 BKT mastery + 매 LLM empathic response + 매 self-report fallback + 매 productive confusion 의 respect.

## 🔗 Graph
- 부모: [[Education-Technology]]
- 변형: [[AutoTutor]] · [[Emotional-AI (Affective Computing)]]
- 응용: [[Adaptive-Learning]]
- Adjacent: [[Emotional-AI (Affective Computing)|Empathy-in-AI]] · [[Dynamic Difficulty Adjustment (DDA)]] · [[Corporate-LMS-Training]]

## 🤖 LLM 활용
**언제**: 매 tutoring product. 매 K-12. 매 language learning.
**언제 X**: 매 reference material. 매 assessment-only.

## ❌ 안티패턴
- **Surveillance feel**: 매 student 의 creep.
- **All confusion = bad**: 매 productive 의 ignore.
- **Static empathy**: 매 personalize X.
- **Privacy violation**: 매 video 의 cloud send.
- **Validate-only**: 매 challenge X.

## 🧪 검증 / 중복
- Verified (D'Mello & Graesser, AutoTutor, 2014+ EITS literature).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-20 | Auto-reinforced |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — affect-aware + 매 detect / strategy / BKT / LLM / A/B code |