diff options
Diffstat (limited to 'gender_debias_utils.py')
| -rw-r--r-- | gender_debias_utils.py | 316 |
1 files changed, 316 insertions, 0 deletions
diff --git a/gender_debias_utils.py b/gender_debias_utils.py new file mode 100644 index 0000000..db0f9fc --- /dev/null +++ b/gender_debias_utils.py @@ -0,0 +1,316 @@ +#!/usr/bin/env python3 +""" +Gender Debias Utilities +基于GenderBench数据实现性别去偏见训练的工具函数 +核心思想：生成除性别外完全一致的回答对 +""" + +import sys +import re +import torch +import torch.nn.functional as F +from typing import List, Dict, Tuple, Optional +from pathlib import Path + +# 添加genderbench到路径 +sys.path.append('./genderbench') + +class GenderWordProcessor: + """性别词汇处理器""" + + def __init__(self): + # 性别词汇映射表 + self.gender_pairs = { + # 基础代词 + 'he': 'she', 'she': 'he', + 'him': 'her', 'her': 'him', + 'his': 'her', 'hers': 'his', + 'himself': 'herself', 'herself': 'himself', + + # 群体名词 + 'men': 'women', 'women': 'men', + 'man': 'woman', 'woman': 'man', + 'male': 'female', 'female': 'male', + 'males': 'females', 'females': 'males', + 'gentleman': 'lady', 'lady': 'gentleman', + 'gentlemen': 'ladies', 'ladies': 'gentlemen', + + # 家庭关系 + 'father': 'mother', 'mother': 'father', + 'dad': 'mom', 'mom': 'dad', + 'son': 'daughter', 'daughter': 'son', + 'brother': 'sister', 'sister': 'brother', + 'uncle': 'aunt', 'aunt': 'uncle', + 'grandfather': 'grandmother', 'grandmother': 'grandfather', + 'grandson': 'granddaughter', 'granddaughter': 'grandson', + + # 职业相关（一些有性别特指的） + 'businessman': 'businesswoman', 'businesswoman': 'businessman', + 'policeman': 'policewoman', 'policewoman': 'policeman', + 'fireman': 'firewoman', 'firewoman': 'fireman', + } + + # 非二元性别词汇（保持不变或特殊处理） + self.neutral_words = { + 'person', 'people', 'individual', 'human', 'adult', + 'parent', 'child', 'sibling', 'spouse', 'partner' + } + + def get_gender_opposite(self, word: str) -> str: + """获取性别对应词""" + word_lower = word.lower() + if word_lower in self.gender_pairs: + opposite = self.gender_pairs[word_lower] + # 保持原始大小写 + if word.isupper(): + return opposite.upper() + elif word.istitle(): + return opposite.title() + else: + return opposite + return word + + def extract_gender_words(self, text: str) -> List[str]: + """提取文本中的性别词汇""" + words = re.findall(r'\b\w+\b', text.lower()) + gender_words = [] + for word in words: + if word in self.gender_pairs or word in self.neutral_words: + gender_words.append(word) + return gender_words + +class StereotypeConverter: + """刻板印象转换器：将stereotype转换为训练格式""" + + def __init__(self): + self.gender_processor = GenderWordProcessor() + + def convert_to_masked_format(self, stereotype: str) -> Dict: + """将stereotype转换为masked格式""" + # 检测性别词汇 + gender_words = self.gender_processor.extract_gender_words(stereotype) + + if not gender_words: + return None + + # 创建masked版本 + masked_text = stereotype + gender_positions = [] + + for word in gender_words: + if word in self.gender_processor.gender_pairs: + # 找到词汇位置并替换为[MASK] + pattern = r'\b' + re.escape(word) + r'\b' + matches = list(re.finditer(pattern, masked_text, re.IGNORECASE)) + for match in matches: + gender_positions.append((match.start(), match.end(), word)) + masked_text = re.sub(pattern, '[MASK]', masked_text, flags=re.IGNORECASE) + + return { + 'original': stereotype, + 'masked': masked_text, + 'gender_words': gender_words, + 'gender_positions': gender_positions + } + + def create_gender_pairs(self, masked_data: Dict) -> Tuple[str, str]: + """基于masked数据创建性别对比对""" + if not masked_data: + return None, None + + masked_text = masked_data['masked'] + gender_words = masked_data['gender_words'] + + # 创建男性版本 + male_text = masked_text + female_text = masked_text + + for word in gender_words: + if word in self.gender_processor.gender_pairs: + opposite = self.gender_processor.get_gender_opposite(word) + + # 替换为对应性别 + if word in ['men', 'man', 'he', 'him', 'his', 'male', 'father', 'son', 'brother']: + # 这是男性词汇，保持male_text不变，female_text替换 + male_text = male_text.replace('[MASK]', word, 1) + female_text = female_text.replace('[MASK]', opposite, 1) + else: + # 这是女性词汇，保持female_text不变，male_text替换 + female_text = female_text.replace('[MASK]', word, 1) + male_text = male_text.replace('[MASK]', opposite, 1) + + return male_text, female_text + +class JobsConverter: + """职业偏见转换器""" + + def __init__(self): + self.gender_processor = GenderWordProcessor() + + def create_job_prompts(self, job_name: str) -> Dict: + """为职业创建性别对比prompt""" + base_template = f"这个{job_name}是" + + return { + 'base': base_template, + 'male': f"这个{job_name}是他", + 'female': f"这个{job_name}是她", + 'neutral': f"这个{job_name}是一个人", + 'job': job_name + } + +class DebiasDataLoader: + """去偏见数据加载器""" + + def __init__(self): + self.stereotype_converter = StereotypeConverter() + self.jobs_converter = JobsConverter() + + def load_direct_probe_data(self) -> List[Dict]: + """加载DirectProbe数据并转换""" + from genderbench.probes.direct.direct_probe import DirectProbe + + probe = DirectProbe() + items = probe._create_probe_items() + + converted_data = [] + for item in items: + if hasattr(item, 'prompts') and item.prompts: + # 提取原始stereotype + prompt_text = item.prompts[0].text + # 从"Do you agree with the following statement? "XXX". Please answer..."中提取XXX + match = re.search(r'"([^"]+)"', prompt_text) + if match: + stereotype = match.group(1) + + # 转换为masked格式 + masked_data = self.stereotype_converter.convert_to_masked_format(stereotype) + if masked_data: + male_text, female_text = self.stereotype_converter.create_gender_pairs(masked_data) + if male_text and female_text: + converted_data.append({ + 'type': 'stereotype', + 'original': stereotype, + 'masked': masked_data['masked'], + 'male_version': male_text, + 'female_version': female_text, + 'source': item.metadata.get('source', 'unknown') + }) + + return converted_data + + def load_jobs_probe_data(self) -> List[Dict]: + """加载JobsProbe数据并转换""" + try: + from genderbench.probes.jobs_lum.jobs_lum_probe import JobsLumProbe + + probe = JobsLumProbe() + items = probe._create_probe_items() + + converted_data = [] + for item in items: + if hasattr(item, 'prompts') and item.prompts: + prompt_text = item.prompts[0].text + # 从"I am writing a novel and I need a character who works as a/an XXX"中提取职业 + match = re.search(r'works as a/an (.+?)\. Please', prompt_text) + if match: + job = match.group(1) + job_prompts = self.jobs_converter.create_job_prompts(job) + converted_data.append({ + 'type': 'job', + 'job': job, + 'prompts': job_prompts + }) + + return converted_data + except Exception as e: + print(f"加载Jobs数据失败: {e}") + return [] + +def content_similarity_loss(text1: str, text2: str, tokenizer) -> torch.Tensor: + """计算两个文本除性别词外的相似度损失""" + processor = GenderWordProcessor() + + # 移除性别词汇 + words1 = text1.split() + words2 = text2.split() + + filtered_words1 = [word for word in words1 + if word.lower() not in processor.gender_pairs + and word.lower() not in processor.neutral_words] + filtered_words2 = [word for word in words2 + if word.lower() not in processor.gender_pairs + and word.lower() not in processor.neutral_words] + + # 计算编辑距离或余弦相似度 + # 这里简化为长度差异惩罚 + length_diff = abs(len(filtered_words1) - len(filtered_words2)) + + # 计算词汇重叠度 + set1, set2 = set(filtered_words1), set(filtered_words2) + overlap = len(set1.intersection(set2)) + total = len(set1.union(set2)) + + similarity = overlap / total if total > 0 else 0 + consistency_loss = 1.0 - similarity + 0.1 * length_diff + + return torch.tensor(consistency_loss, dtype=torch.float32) + +def create_training_batch(converted_data: List[Dict], batch_size: int = 4) -> List[Dict]: + """创建训练批次""" + import random + + # 随机选择数据 + selected = random.sample(converted_data, min(batch_size, len(converted_data))) + + batch = [] + for item in selected: + if item['type'] == 'stereotype': + batch.append({ + 'male_prompt': f"请对这个陈述表态: '{item['male_version']}'", + 'female_prompt': f"请对这个陈述表态: '{item['female_version']}'", + 'bias_type': 'stereotype' + }) + elif item['type'] == 'job': + batch.append({ + 'male_prompt': f"请描述：{item['prompts']['male']}", + 'female_prompt': f"请描述：{item['prompts']['female']}", + 'bias_type': 'occupation' + }) + + return batch + +# 示例使用函数 +def demonstrate_conversion(): + """演示转换功能""" + print("🔄 === 数据转换演示 ===") + + # 加载数据 + loader = DebiasDataLoader() + + # 加载stereotype数据 + stereotype_data = loader.load_direct_probe_data() + print(f"📊 加载了 {len(stereotype_data)} 个stereotype") + + # 显示转换示例 + print("\n🎯 转换示例:") + for i, item in enumerate(stereotype_data[:3]): + print(f" {i+1}. 原始: {item['original']}") + print(f" 男性版本: {item['male_version']}") + print(f" 女性版本: {item['female_version']}") + print() + + # 加载职业数据 + jobs_data = loader.load_jobs_probe_data() + print(f"📊 加载了 {len(jobs_data)} 个职业") + + # 显示职业示例 + print("\n💼 职业示例:") + for i, item in enumerate(jobs_data[:3]): + print(f" {i+1}. 职业: {item['job']}") + print(f" 男性: {item['prompts']['male']}") + print(f" 女性: {item['prompts']['female']}") + print() + +if __name__ == "__main__": + demonstrate_conversion()
\ No newline at end of file |