---
id: wiki-2026-0508-emotional-ai-affective-computing
title: Emotional AI (Affective Computing)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [affective computing, emotion AI, sentiment analysis, emotion recognition, FER]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [ai, affective-computing, emotional-ai, sentiment, fer, multimodal, hci]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: PyTorch / Transformers / DeepFace
---

# Emotional AI (Affective Computing)

## 매 한 줄
> **"매 emotion 의 sense + 매 generate"**. Picard MIT (1997). 매 facial, voice, text, physiological. 매 modern: 매 multimodal LLM + valence-arousal regression. 매 ethical: 매 surveillance + cultural bias.

## 매 핵심

### 매 emotion model
- **Discrete** (Ekman): 매 6 basic — anger, fear, disgust, surprise, sadness, joy.
- **Dimensional** (Russell): 매 valence × arousal.
- **Plutchik wheel**: 매 8 + intensity.
- **Appraisal** (Scherer): 매 cognitive evaluation.

### 매 modality
- **Facial** (FER): 매 AU (action unit).
- **Voice**: 매 prosody, pitch.
- **Text**: 매 lexicon + transformer.
- **Physiological**: 매 HRV, GSR, EEG.
- **Body**: 매 posture, gait.
- **Multimodal**: 매 fuse.

### 매 응용
1. **Customer service**: 매 sentiment.
2. **Mental health**: 매 mood track.
3. **Education**: 매 engagement.
4. **Auto**: 매 driver drowsiness.
5. **Marketing**: 매 ad reaction.
6. **Robot**: 매 emotional support.
7. **Game**: 매 dynamic difficulty.

### 매 ethical
- **Privacy**: 매 emotion 의 surveil.
- **Cultural bias**: 매 Ekman universality 의 contested.
- **Inaccuracy**: 매 expression ≠ feeling.
- **Manipulation**: 매 ad / dark pattern.
- **Regulation**: 매 EU AI Act 의 emotion 의 high-risk.

## 💻 패턴

### Text sentiment (transformers)
```python
from transformers import pipeline
pipe = pipeline('sentiment-analysis', model='cardiffnlp/twitter-roberta-base-sentiment')
print(pipe("I love this!"))
# 매 [{'label': 'POSITIVE', 'score': 0.99}]
```

### Emotion classification (text, 7-class)
```python
classifier = pipeline('text-classification', model='j-hartmann/emotion-english-distilroberta-base', top_k=None)
result = classifier("I'm worried about tomorrow")
# 매 [{'label': 'fear', 'score': 0.7}, ...]
```

### Facial emotion (DeepFace)
```python
from deepface import DeepFace
result = DeepFace.analyze('selfie.jpg', actions=['emotion'])
print(result[0]['dominant_emotion'])  # 매 'happy'
```

### Voice emotion (audio + Wav2Vec2)
```python
import librosa
import torch
from transformers import Wav2Vec2ForSequenceClassification, Wav2Vec2FeatureExtractor

model = Wav2Vec2ForSequenceClassification.from_pretrained('superb/wav2vec2-base-superb-er')
fe = Wav2Vec2FeatureExtractor.from_pretrained('superb/wav2vec2-base-superb-er')

def predict_emotion(audio_path):
    audio, sr = librosa.load(audio_path, sr=16000)
    inputs = fe(audio, sampling_rate=16000, return_tensors='pt')
    with torch.no_grad():
        logits = model(**inputs).logits
    return model.config.id2label[logits.argmax().item()]
```

### Valence-Arousal regression
```python
class VARegressor(nn.Module):
    def __init__(self, backbone):
        super().__init__()
        self.backbone = backbone
        self.va_head = nn.Linear(backbone.hidden_dim, 2)  # 매 valence, arousal
    
    def forward(self, x):
        feat = self.backbone(x)
        va = torch.tanh(self.va_head(feat))  # 매 [-1, 1]
        return va  # 매 [batch, (valence, arousal)]
```

### Multimodal fusion
```python
class MultiModalEmotion(nn.Module):
    def __init__(self):
        super().__init__()
        self.face_enc = FacialEncoder()
        self.audio_enc = AudioEncoder()
        self.text_enc = TextEncoder()
        self.fusion = nn.Linear(768 * 3, 7)
    
    def forward(self, face, audio, text):
        ff = self.face_enc(face)
        fa = self.audio_enc(audio)
        ft = self.text_enc(text)
        return self.fusion(torch.cat([ff, fa, ft], dim=-1))
```

### Physiological (HRV stress)
```python
import neurokit2 as nk
def stress_from_ecg(ecg_signal, sr=1000):
    signals, info = nk.ecg_process(ecg_signal, sampling_rate=sr)
    hrv = nk.hrv_time(signals, sampling_rate=sr)
    rmssd = hrv['HRV_RMSSD'].iloc[0]
    return 'stressed' if rmssd < 30 else 'calm'
```

### Cultural-aware (avoid bias)
```python
def culturally_aware_predict(image, region):
    base_pred = model.predict(image)
    # 매 cultural calibration
    if region == 'east_asia':
        # 매 East Asian 의 less expressive baseline
        base_pred['happy'] *= 0.8
        base_pred['neutral'] *= 1.2
    return normalize(base_pred)
```

### LLM-based emotion (modern)
```python
def llm_emotion(text):
    prompt = f"""Classify the emotion in this text. Output JSON.
Text: "{text}"
Output: {{"primary": "...", "secondary": "...", "valence": -1 to 1, "arousal": 0 to 1}}"""
    return json.loads(llm.generate(prompt))
```

### Real-time engagement (online learning)
```python
class EngagementTracker:
    def __init__(self, baseline=0.5):
        self.engagement = baseline
        self.history = []
    
    def update(self, frame):
        emotion = analyze_face(frame)
        self.engagement = 0.9 * self.engagement + 0.1 * emotion['attention_score']
        self.history.append(self.engagement)
    
    def is_disengaged(self):
        return np.mean(self.history[-30:]) < 0.3
```

### Affective response (chatbot)
```python
def empathic_response(user_msg):
    emotion = classify(user_msg)
    if emotion in ('sad', 'fear'):
        return llm.generate(f"User feels {emotion}. Respond with validation first, then gentle reframe.\nUser: {user_msg}")
    return llm.generate(f"Respond to: {user_msg}")
```

### Privacy-aware (on-device)
```python
# 매 raw frame 의 server 의 send X
def on_device_emotion(frame, local_model):
    emotion = local_model(frame)
    # 매 only summary 의 send
    return {'emotion': emotion, 'confidence': ...}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Customer support text | Sentiment + LLM |
| Driver drowsiness | Facial + on-device |
| Mental health | Multimodal + clinician |
| Marketing | A/B + reaction |
| Robot pet | Multimodal real-time |
| Education | Engagement (eye + face) |

**기본값**: 매 task-specific modality + 매 cultural calibration + 매 privacy on-device + 매 LLM augment + 매 ethical disclosure.

## 🔗 Graph
- 부모: [[AI]] · [[Human Computer Interaction]]
- 변형: [[Sentiment-Analysis]] · [[FER]] · [[Emotional-AI (Affective Computing)|Empathy-in-AI]]
- Adjacent: [[Multimodal-LLM]] · [[Drama Management Systems]] · [[Dynamic Difficulty Adjustment (DDA)]] · [[EU-AI-Act]]

## 🤖 LLM 활용
**언제**: 매 customer experience. 매 health (with clinician). 매 robot interaction.
**언제 X**: 매 surveillance. 매 manipulation. 매 EU high-risk.

## ❌ 안티패턴
- **Single-modality trust**: 매 noisy.
- **Universal Ekman 의 assume**: 매 cultural bias.
- **Express ≠ feel**: 매 mask.
- **No consent**: 매 surveillance.
- **Manipulate emotion**: 매 dark pattern.

## 🧪 검증 / 중복
- Verified (Picard 1997, Russell circumplex, Barrett constructionist).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-04-26 | EMOTION-AI auto |
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — emotion model + 매 text/voice/face/VA/multimodal code |