src/common/constants.py

# 공용 상수

EMOTION_LABELS = [
    "neutral",
    "joy",
    "sadness",
    "anger",
    "surprise",
    "fear",
    "disgust",
]

EMOTION_TO_VALENCE = {
    "joy": 1.0,
    "surprise": 0.5,
    "neutral": 0.0,
    "sadness": -0.5,
    "fear": -0.6,
    "anger": -0.8,
    "disgust": -0.9,
}

FUSION_WEIGHTS = {
    "audio": 0.6,
    "text": 0.4,
}

# Korean emotion-specific fusion weights — trained via gradient descent
# Parameterization: w_a = sigmoid(α), w_t = 1 - w_a (7 params jointly optimized)
# Data: AI Hub 263 val (1,294 samples, stratified 80/20 train/val)
# Val macro F1 = 0.8724 (vs 0.8744 fixed 60/40, 0.8757 greedy v1).
# Full 263 val macro F1 = 0.8748 (vs 0.8736 greedy v1, 0.8347 fixed 60/40).
# Text-dominant pattern preserved (Korean KcELECTRA LoRA > audio LoRA on most classes).
EMOTION_FUSION_WEIGHTS_KO = {
    "neutral":  {"audio": 0.53, "text": 0.47},
    "joy":      {"audio": 0.25, "text": 0.75},
    "sadness":  {"audio": 0.16, "text": 0.84},
    "anger":    {"audio": 0.11, "text": 0.89},
    "surprise": {"audio": 0.22, "text": 0.78},
    "fear":     {"audio": 0.10, "text": 0.90},
    "disgust":  {"audio": 0.15, "text": 0.85},
}

# English emotion-specific fusion weights — trained via gradient descent
# Parameterization: w_a = sigmoid(α), w_t = 1 - w_a (7 params jointly optimized)
# Data: JL-Corpus + SAVEE + MELD + RAVDESS phone (fear/disgust/sadness) = 2,821 samples
# Val macro F1 = 0.7596 (vs 0.7473 fixed 60/40, 0.7114 audio-only, 0.6295 greedy v1)
EMOTION_FUSION_WEIGHTS_EN = {
    "neutral":  {"audio": 0.78, "text": 0.22},
    "joy":      {"audio": 0.61, "text": 0.39},
    "sadness":  {"audio": 0.70, "text": 0.30},
    "anger":    {"audio": 0.60, "text": 0.40},
    "surprise": {"audio": 0.46, "text": 0.54},
    "fear":     {"audio": 0.74, "text": 0.26},
    "disgust":  {"audio": 0.77, "text": 0.23},
}

# Default (Korean) — back-compat alias
EMOTION_FUSION_WEIGHTS = EMOTION_FUSION_WEIGHTS_KO