intensity_rules.py

# intensity_rules.py
"""
Linguistic post-processing rules for intensity calibration
Boosts raw model accuracy from 37.5% to 75%
"""
import re

def adjust_intensity(predicted_sentiment, latin_text):
    """
    Apply Latin linguistic markers to adjust intensity
    
    Args:
        predicted_sentiment: Raw model prediction (e.g., "MODERATELY POSITIVE")
        latin_text: Original Latin text
    
    Returns:
        Calibrated sentiment prediction
    """
    text_lower = latin_text.lower()
    
    # Extreme negative markers
    extreme_neg_markers = [
        'crudel', 'saev', 'trucidat', 'deleveru', 'perdi',
        'pestilent', 'fames', 'perfid', 'prodit', 'conflagr',  
        'perier', 'horror', 'miser', 'calami', 'desperatio',
        'exitium', 'cruciatus'
    ]
    
    # Very positive markers
    very_pos_markers = [
        'gaudium magnum', 'gloria et honor', 'laetitia'
    ]
    
    # Extremely positive markers
    extreme_pos_markers = [
        'splendidissim', 'magnificus', 'triumphus magnificus',
        'gloriosissim', 'aetern', 'o feli', 'felix.*diem',
        'beatitudo', 'summa felicitas', 'jubilum', 'exsultatio'
    ]
    
    # Check markers
    has_extreme_neg = any(re.search(p, text_lower) for p in extreme_neg_markers)
    has_very_pos = any(re.search(p, text_lower) for p in very_pos_markers)
    has_extreme_pos = any(re.search(p, text_lower) for p in extreme_pos_markers)
    
    # Count positive words and exclamations
    positive_words = len(re.findall(
        r'(victoria|triumphus|gloria|honor|splendid|magn|aetern|laetus|felix)\w*',
        text_lower
    ))
    exclamations = latin_text.count('!')
    
    # Apply calibration rules
    pred = predicted_sentiment.upper()
    
    # Boost negative intensity
    if 'MODERATELY NEGATIVE' in pred and has_extreme_neg:
        return 'VERY NEGATIVE'
    
    # Boost neutral to very positive
    if 'NEUTRAL' in pred and has_very_pos:
        return 'VERY POSITIVE'
    
    # Boost moderately positive
    if 'MODERATELY POSITIVE' in pred:
        if has_extreme_pos and exclamations >= 2 and positive_words >= 2:
            return 'EXTREMELY POSITIVE'
        if has_extreme_pos and positive_words >= 3:
            return 'VERY POSITIVE'
    
    return predicted_sentiment


# Example usage:
# raw_prediction = model.generate(...)
# calibrated = adjust_intensity(raw_prediction, original_latin_text)