evaluation/debug_failures.py

# Quick debug script to analyze failing emotions
import sys
sys.path.insert(0, '.')

from avatar.sentiment_multi_emotion import MultiEmotionAnalyzer

a = MultiEmotionAnalyzer()

tests = [
    ('I feel so calm and peaceful', 'calm', 'positive'),
    ('I feel such affection for you', 'affection', 'positive'),
    ('Such tender feelings', 'tenderness', 'positive'),
    ('I feel compassion for them', 'compassion', 'positive'),
    ('I empathize with you', 'empathy', 'positive'),
    ('What a fascinating discovery!', 'fascination', 'positive'),
    ('I stand in awe of this', 'awe', 'positive'),
    ('This is intriguing', 'intrigue', 'positive'),
    ('I am in shock!', 'shock', 'neutral'),
    ('I am confused about this', 'confused', 'neutral'),
    ('I am puzzled by this', 'puzzled', 'neutral'),
]

print("=" * 70)
print("FAILING EMOTION ANALYSIS")
print("=" * 70)
print()

for text, emotion, expected in tests:
    r = a.analyze(text)
    status = "✓" if r['polarity'] == expected else "✗"
    print(f"{status} {emotion:15} | expected: {expected:8} | got: {r['polarity']:8} | base: {r['base_emotion']:8} | label: {r['label']}")