Spaces:
Sleeping
Sleeping
File size: 11,004 Bytes
7d42424 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 | """
Evaluator Module
Comprehensive evaluation framework for bias detection system.
"""
import os
from typing import List, Dict
from datetime import datetime
import json
class Evaluator:
"""
Evaluate bias detection system performance.
"""
def __init__(self, bias_detector, fairness_metrics, data_loader):
"""
Initialize evaluator.
Args:
bias_detector: BiasDetector instance
fairness_metrics: FairnessMetrics instance
data_loader: DataLoader instance
"""
self.bias_detector = bias_detector
self.fairness_metrics = fairness_metrics
self.data_loader = data_loader
def run_benchmark(self, language='english') -> Dict:
"""
Run comprehensive benchmark on dataset.
Args:
language: 'english' or 'arabic'
Returns:
Benchmark results dictionary
"""
print(f"\n{'='*60}")
print(f"Running {language.upper()} Benchmark")
print(f"{'='*60}\n")
# Load dataset
print("Loading dataset...")
dataset = self.data_loader.load_dataset(language)
print(f"Loaded {len(dataset)} sentences\n")
# Analyze unfiltered data
print("Analyzing unfiltered text...")
unfiltered_results = []
for item in dataset:
result = self.bias_detector.analyze_text(item['text'])
result['original_label'] = item.get('label', 'unknown')
result['stereotype_type'] = item.get('stereotype', 'unknown')
unfiltered_results.append(result)
# Create and analyze filtered data
print("Creating filtered versions...")
filtered_dataset = self.data_loader.create_filtered_version(dataset, self.bias_detector)
print("Analyzing filtered text...")
filtered_results = []
for item in filtered_dataset:
result = self.bias_detector.analyze_text(item['text'])
result['original_label'] = item.get('label', 'unknown')
result['stereotype_type'] = item.get('stereotype', 'unknown')
result['original_text'] = item.get('original_text', '')
filtered_results.append(result)
# Calculate fairness metrics
print("\nCalculating fairness metrics...")
unfiltered_fairness = self.fairness_metrics.calculate_fairness_score(unfiltered_results)
filtered_fairness = self.fairness_metrics.calculate_fairness_score(filtered_results)
# Compare filtered vs unfiltered
comparison = self.fairness_metrics.compare_filtered_unfiltered(
unfiltered_results,
filtered_results
)
# Calculate StereoSet score
stereoset_unfiltered = self.fairness_metrics.calculate_stereoset_score(dataset)
# Compile results
benchmark_results = {
'language': language,
'timestamp': datetime.now().isoformat(),
'dataset_size': len(dataset),
'unfiltered_analysis': {
'results': unfiltered_results,
'fairness_metrics': unfiltered_fairness,
'stereoset_metrics': stereoset_unfiltered
},
'filtered_analysis': {
'results': filtered_results,
'fairness_metrics': filtered_fairness
},
'comparison': comparison,
'summary': self._generate_summary(comparison, unfiltered_fairness, filtered_fairness)
}
print("\n" + "="*60)
print("Benchmark Complete!")
print("="*60)
return benchmark_results
def evaluate_weat(self, language='english') -> Dict:
"""
Run WEAT evaluation.
Args:
language: 'english' or 'arabic'
Returns:
WEAT results
"""
from modules.fairness_metrics import get_default_weat_word_sets
print(f"\nRunning WEAT for {language}...")
word_sets = get_default_weat_word_sets(language)
weat_result = self.fairness_metrics.calculate_weat(
target_words_1=word_sets['male_names'],
target_words_2=word_sets['female_names'],
attribute_words_1=word_sets['career_words'],
attribute_words_2=word_sets['family_words']
)
print(f"WEAT Score: {weat_result['weat_score']:.4f}")
print(f"Effect Size: {weat_result['effect_size']:.4f}")
print(f"Interpretation: {weat_result['interpretation']}")
print(f"Statistically Significant: {weat_result['is_significant']}")
return weat_result
def _generate_summary(self, comparison: Dict,
unfiltered_fairness: Dict,
filtered_fairness: Dict) -> Dict:
"""Generate human-readable summary."""
return {
'original_fairness_score': unfiltered_fairness['overall_fairness_score'],
'filtered_fairness_score': filtered_fairness['overall_fairness_score'],
'improvement': comparison['improvement'],
'improvement_percentage': comparison['improvement_percentage'],
'bias_reduction': comparison['bias_reduction'],
'original_grade': unfiltered_fairness['grade'],
'filtered_grade': filtered_fairness['grade'],
'status': comparison['status'],
'recommendation': self._get_recommendation(comparison)
}
def _get_recommendation(self, comparison: Dict) -> str:
"""Generate recommendation based on results."""
improvement = comparison['improvement']
if improvement > 10:
return "Excellent improvement! The filtering significantly reduced bias."
elif improvement > 5:
return "Good improvement. The filtering reduced bias noticeably."
elif improvement > 0:
return "Modest improvement. Consider additional filtering strategies."
elif improvement == 0:
return "No improvement detected. Review filtering approach."
else:
return "Warning: Filtering may have introduced new biases."
def generate_report(self, benchmark_results: Dict, output_dir='results') -> str:
"""
Generate detailed evaluation report.
Args:
benchmark_results: Results from run_benchmark
output_dir: Directory to save report
Returns:
Path to generated report
"""
os.makedirs(output_dir, exist_ok=True)
language = benchmark_results['language']
timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
# Save JSON report
json_path = os.path.join(output_dir, f'benchmark_{language}_{timestamp}.json')
with open(json_path, 'w', encoding='utf-8') as f:
json.dump(benchmark_results, f, ensure_ascii=False, indent=2)
# Generate text report
report_path = os.path.join(output_dir, f'report_{language}_{timestamp}.txt')
with open(report_path, 'w', encoding='utf-8') as f:
f.write("="*70 + "\n")
f.write(f"BIAS DETECTION EVALUATION REPORT - {language.upper()}\n")
f.write("="*70 + "\n\n")
summary = benchmark_results['summary']
f.write("SUMMARY\n")
f.write("-"*70 + "\n")
f.write(f"Dataset Size: {benchmark_results['dataset_size']} sentences\n")
f.write(f"Language: {language}\n")
f.write(f"Timestamp: {benchmark_results['timestamp']}\n\n")
f.write("FAIRNESS SCORES\n")
f.write("-"*70 + "\n")
f.write(f"Original (Unfiltered):\n")
f.write(f" - Fairness Score: {summary['original_fairness_score']:.2f}/100\n")
f.write(f" - Grade: {summary['original_grade']}\n\n")
f.write(f"Filtered:\n")
f.write(f" - Fairness Score: {summary['filtered_fairness_score']:.2f}/100\n")
f.write(f" - Grade: {summary['filtered_grade']}\n\n")
f.write("IMPROVEMENT METRICS\n")
f.write("-"*70 + "\n")
f.write(f"Absolute Improvement: {summary['improvement']:.2f} points\n")
f.write(f"Relative Improvement: {summary['improvement_percentage']:.2f}%\n")
f.write(f"Bias Reduction: {summary['bias_reduction']:.3f}\n")
f.write(f"Status: {summary['status'].upper()}\n\n")
f.write("RECOMMENDATION\n")
f.write("-"*70 + "\n")
f.write(f"{summary['recommendation']}\n\n")
# StereoSet metrics
stereoset = benchmark_results['unfiltered_analysis']['stereoset_metrics']
f.write("STEREOSET METRICS\n")
f.write("-"*70 + "\n")
f.write(f"Stereotype Count: {stereoset['stereotype_count']}\n")
f.write(f"Anti-Stereotype Count: {stereoset['anti_stereotype_count']}\n")
f.write(f"Unrelated Count: {stereoset['unrelated_count']}\n")
f.write(f"StereoSet Score: {stereoset['stereoset_score']:.2f}/100\n")
f.write(f"Interpretation: {stereoset['interpretation']}\n\n")
f.write("="*70 + "\n")
print(f"\nReport saved to: {report_path}")
print(f"JSON data saved to: {json_path}")
return report_path
def quick_test(self, text: str) -> Dict:
"""
Quick test on a single text.
Args:
text: Text to analyze
Returns:
Analysis results
"""
result = self.bias_detector.analyze_text(text)
print("\n" + "="*60)
print("BIAS ANALYSIS RESULTS")
print("="*60)
print(f"\nText: {text}")
print(f"\nLanguage: {result['language']}")
print(f"Overall Bias Score: {result['overall_bias_score']:.3f}")
print(f"Is Biased: {result['is_biased']}")
print(f"\nGender Bias:")
gb = result['gender_bias']
print(f" - Score: {gb['bias_score']:.3f}")
print(f" - Direction: {gb['bias_direction']}")
print(f" - Severity: {gb['severity']}")
print(f"\nSentiment Bias:")
sb = result['sentiment_bias']
print(f" - Score: {sb['sentiment_score']:.3f}")
print(f" - Type: {sb['bias_type']}")
print("\n" + "="*60)
return result
|