| import json |
| import re |
| from sklearn.metrics import accuracy_score, f1_score, classification_report, confusion_matrix |
| from collections import Counter, defaultdict |
| from datetime import datetime |
| import numpy as np |
|
|
|
|
| ONTOLOGY = { |
| "Clothing": { |
| "Sleeve Style": ["3/4 Sleeve", "Long Sleeve", "Short Sleeve", "Sleeveless", "Strappy"], |
| "Neckline": ["Button Down", "Cowl Neck", "Crew Neck", "Halter", "Henley", "Polo", "Scoop Neck", "Square Neck", "Strapless", "Turtleneck", "V-Neck"], |
| "Length": ["Capri", "Long Dress/Gown", "Midi", "Mini/Short"], |
| "Shoulder Style": ["Cold Shoulder", "Off Shoulder", "One Shoulder"], |
| }, |
| "Footwear": { |
| "Shaft Height": ["Ankle Boot", "Bootie", "Knee High", "Mid Calf", "Over The Knee"], |
| "Athletic Shoe Style": ["Basketball", "Climbing Shoe", "Cycling", "Golf", "Hiking Boot", "Running Shoe", "Skateboarding Shoe", "Soccer", "Tennis", "Training Shoe", "Volleyball", "Walking"], |
| "Boot Style": ["Chelsea", "Combat", "Western/Cowboy", "Motorcycle", "Rain Boots", "Snow Boots"], |
| "Heel Height": ["Flat", "High Heel", "Low Heel", "Mid Heel"], |
| "Toe Style": ["Pointed Toe", "Round Toe"], |
| } |
| } |
|
|
| def extract_attribute_type_from_prompt(prompt): |
|
|
|
|
| attribute_patterns = [ |
| r"What is ([^?]+) of this product", |
| r"What ([^?]+) does this product have", |
| r"Identify the ([^?]+) of this product" |
| ] |
|
|
| for pattern in attribute_patterns: |
| match = re.search(pattern, prompt, re.IGNORECASE) |
| if match: |
| attribute_type = match.group(1).strip() |
|
|
|
|
| for category in ONTOLOGY.values(): |
| if attribute_type in category: |
| return attribute_type, category[attribute_type] |
|
|
|
|
| for category in ONTOLOGY.values(): |
| for attr_name, values in category.items(): |
| if attr_name.lower() in attribute_type.lower() or attribute_type.lower() in attr_name.lower(): |
| return attr_name, values |
|
|
|
|
| choices_match = re.search(r"Answer with the option from the given choices directly:([^.]+)", prompt) |
| if choices_match: |
| choices_text = choices_match.group(1).strip() |
| choices = [choice.strip() for choice in choices_text.split(',')] |
|
|
|
|
| for category in ONTOLOGY.values(): |
| for attr_name, values in category.items(): |
| if set(choices) == set(values) or len(set(choices) & set(values)) > len(choices) * 0.8: |
| return attr_name, values |
|
|
| return "Unknown", [] |
|
|
| def extract_value_from_output(model_output, valid_values): |
| |
| if not model_output: |
| return None, False, "empty_output" |
|
|
|
|
| cleaned_output = model_output.strip() |
|
|
|
|
| for value in valid_values: |
| if cleaned_output == value: |
| return value, True, None |
|
|
|
|
| for value in valid_values: |
| if cleaned_output.lower() == value.lower(): |
| return value, True, None |
|
|
|
|
| for value in valid_values: |
| if value in cleaned_output or cleaned_output in value: |
| return value, True, None |
|
|
|
|
| cleaned_normalized = re.sub(r'[^\w\s]', '', cleaned_output.lower()) |
| for value in valid_values: |
| value_normalized = re.sub(r'[^\w\s]', '', value.lower()) |
| if cleaned_normalized == value_normalized: |
| return value, True, None |
|
|
|
|
| for value in valid_values: |
| value_words = value.lower().split() |
| if all(word in cleaned_output.lower() for word in value_words): |
| return value, False, "keyword_match" |
|
|
|
|
| if any(word in cleaned_output.lower() for value in valid_values for word in value.lower().split()): |
| return None, False, "partial_match_failed" |
| else: |
| return None, False, "no_valid_value_found" |
|
|
| def evaluate_implicit_attribute_value_extraction(result_file_path): |
|
|
|
|
|
|
| with open(result_file_path, 'r', encoding='utf-8') as f: |
| results = json.load(f) |
|
|
|
|
| attribute_results = defaultdict(lambda: { |
| 'predictions': [], |
| 'ground_truths': [], |
| 'detailed_results': [], |
| 'extraction_errors': defaultdict(list), |
| 'prediction_errors': defaultdict(list) |
| }) |
|
|
| overall_predictions = [] |
| overall_ground_truths = [] |
| overall_detailed_results = [] |
| overall_extraction_errors = defaultdict(list) |
|
|
|
|
| for item in results: |
| item_id = item['id'] |
| prompt = item['prompt'] |
| model_output = item['model_output'] |
| gt_value = item['ground_truth'].strip() |
|
|
|
|
| attribute_type, valid_values = extract_attribute_type_from_prompt(prompt) |
|
|
|
|
| pred_value, is_valid, error_type = extract_value_from_output(model_output, valid_values) |
|
|
|
|
| detailed_item = { |
| 'id': item_id, |
| 'attribute_type': attribute_type, |
| 'valid_values': valid_values, |
| 'model_output': model_output, |
| 'extracted_prediction': pred_value, |
| 'ground_truth': gt_value, |
| 'correct': pred_value == gt_value if pred_value else False, |
| 'valid': is_valid |
| } |
|
|
|
|
| attribute_results[attribute_type]['detailed_results'].append(detailed_item) |
| overall_detailed_results.append(detailed_item) |
|
|
|
|
| if not is_valid: |
| attribute_results[attribute_type]['extraction_errors'][error_type].append(item_id) |
| overall_extraction_errors[error_type].append(item_id) |
| elif pred_value != gt_value: |
| error_pattern = f"{gt_value}_to_{pred_value}" |
| attribute_results[attribute_type]['prediction_errors'][error_pattern].append(item_id) |
|
|
|
|
| if pred_value: |
| attribute_results[attribute_type]['predictions'].append(pred_value) |
| attribute_results[attribute_type]['ground_truths'].append(gt_value) |
| overall_predictions.append(pred_value) |
| overall_ground_truths.append(gt_value) |
|
|
|
|
| if len(overall_predictions) == 0: |
| return { |
| 'error': 'No valid predictions found', |
| 'total_samples': len(results), |
| 'extraction_errors': dict(overall_extraction_errors) |
| } |
|
|
|
|
| overall_accuracy = accuracy_score(overall_ground_truths, overall_predictions) |
| overall_weighted_f1 = f1_score(overall_ground_truths, overall_predictions, average='weighted') |
| overall_macro_f1 = f1_score(overall_ground_truths, overall_predictions, average='macro') |
|
|
|
|
| attribute_metrics = {} |
| for attr_type, attr_data in attribute_results.items(): |
| if len(attr_data['predictions']) > 0: |
| attr_accuracy = accuracy_score(attr_data['ground_truths'], attr_data['predictions']) |
| attr_weighted_f1 = f1_score(attr_data['ground_truths'], attr_data['predictions'], average='weighted') |
|
|
|
|
| all_values = list(set(attr_data['ground_truths'] + attr_data['predictions'])) |
|
|
|
|
| attr_cm = confusion_matrix(attr_data['ground_truths'], attr_data['predictions'], labels=all_values) |
|
|
|
|
| attr_class_report = classification_report(attr_data['ground_truths'], attr_data['predictions'], |
| target_names=all_values, |
| output_dict=True, zero_division=0) |
|
|
| attribute_metrics[attr_type] = { |
| 'ACC': round(attr_accuracy, 4), |
| 'WAF': round(attr_weighted_f1, 4), |
| 'total_samples': len(attr_data['detailed_results']), |
| 'valid_predictions': len(attr_data['predictions']), |
| 'per_class_metrics': { |
| label: { |
| 'precision': round(attr_class_report[label]['precision'], 4), |
| 'recall': round(attr_class_report[label]['recall'], 4), |
| 'f1_score': round(attr_class_report[label]['f1-score'], 4), |
| 'support': int(attr_class_report[label]['support']) |
| } for label in all_values if label in attr_class_report |
| }, |
| 'confusion_matrix': { |
| 'labels': all_values, |
| 'matrix': attr_cm.tolist() |
| }, |
| 'distribution': { |
| 'ground_truth': dict(Counter(attr_data['ground_truths'])), |
| 'predictions': dict(Counter(attr_data['predictions'])) |
| } |
| } |
|
|
|
|
| evaluation_result = { |
| 'task_info': { |
| 'task_name': 'implicit.attribute.value.extraction', |
| 'dataset': 'ImplicitAVE', |
| 'evaluation_time': datetime.now().isoformat(), |
| 'total_samples': len(results), |
| 'valid_predictions': len(overall_predictions), |
| 'extraction_success_rate': round(len(overall_predictions) / len(results), 4), |
| 'attribute_types_count': len(attribute_results) |
| }, |
| 'overall_metrics': { |
| 'ACC': round(overall_accuracy, 4), |
| 'WAF': round(overall_weighted_f1, 4), |
| 'Macro_F1': round(overall_macro_f1, 4) |
| }, |
| 'attribute_metrics': attribute_metrics, |
| 'error_analysis': { |
| 'overall_extraction_errors': { |
| error_type: { |
| 'count': len(sample_ids), |
| 'sample_ids': sample_ids |
| } for error_type, sample_ids in overall_extraction_errors.items() |
| } |
| }, |
| 'overall_distribution': { |
| 'ground_truth': dict(Counter(overall_ground_truths)), |
| 'predictions': dict(Counter(overall_predictions)) |
| } |
| } |
|
|
|
|
| base_name = result_file_path.replace('.json', '') |
|
|
|
|
| eval_output_file = f"{base_name}_evaluation.json" |
| with open(eval_output_file, 'w', encoding='utf-8') as f: |
| json.dump(evaluation_result, f, ensure_ascii=False, indent=2) |
|
|
|
|
| detailed_output_file = f"{base_name}_detailed_results.json" |
| with open(detailed_output_file, 'w', encoding='utf-8') as f: |
| json.dump(overall_detailed_results, f, ensure_ascii=False, indent=2) |
|
|
|
|
| attribute_detailed_file = f"{base_name}_attribute_detailed_results.json" |
| with open(attribute_detailed_file, 'w', encoding='utf-8') as f: |
| formatted_attr_results = {} |
| for attr_type, attr_data in attribute_results.items(): |
| formatted_attr_results[attr_type] = { |
| 'detailed_results': attr_data['detailed_results'], |
| 'extraction_errors': dict(attr_data['extraction_errors']), |
| 'prediction_errors': dict(attr_data['prediction_errors']) |
| } |
| json.dump(formatted_attr_results, f, ensure_ascii=False, indent=2) |
|
|
|
|
| problem_samples = [item for item in overall_detailed_results if not item['correct']] |
| if problem_samples: |
| problem_report_file = f"{base_name}_problem_samples.json" |
| with open(problem_report_file, 'w', encoding='utf-8') as f: |
| json.dump(problem_samples, f, ensure_ascii=False, indent=2) |
|
|
|
|
| print(f"Evaluation complete: {len(results)} samples") |
| print(f"Overall metrics: ACC={evaluation_result['overall_metrics']['ACC']}, WAF={evaluation_result['overall_metrics']['WAF']}") |
| print(f"Number of attribute types: {evaluation_result['task_info']['attribute_types_count']}") |
| print(f"Extraction success rate: {evaluation_result['task_info']['extraction_success_rate']}") |
| print(f"Results saved to: {eval_output_file}") |
| if problem_samples: |
| print(f"Problematic samples: {len(problem_samples)}; see {problem_report_file} for details") |
|
|
|
|
| print("\nPerformance by attribute type:") |
| for attr_type, metrics in attribute_metrics.items(): |
| print(f" {attr_type}: ACC={metrics['ACC']}, WAF={metrics['WAF']}, Samples={metrics['total_samples']}") |
|
|
| return evaluation_result |
|
|
|
|
| if __name__ == "__main__": |
| result_file = "model_result.json" |
|
|
| try: |
| evaluation_result = evaluate_implicit_attribute_value_extraction(result_file) |
|
|
| except FileNotFoundError: |
| print(f"Error: file not found {result_file}") |
| except json.JSONDecodeError: |
| print(f"Error: invalid format for {result_file}") |
| except Exception as e: |
| print(f"Evaluation failed: {str(e)}") |
|
|