内容安全与审核规范

def l1_moderate(text, user_id, context): scores = {} # 1. 关键词快速过滤（高召回） keyword_hits = keyword_filter.scan(text) if keyword_hits: scores['keyword'] = max(h.severity for h in keyword_hits) # 2. Perspective API（毒性评分） toxicity = perspective_api.score(text, attributes=['TOXICITY','INSULT','THREAT']) scores['toxicity'] = toxicity['TOXICITY'] # 3. 自研分类模型 classification = bert_classifier.predict(text) scores['classifier'] = classification.max_prob # 4. 用户信誉加权 trust = get_user_trust_score(user_id) # 0-100 weight = 1.0 + (50 - trust) / 100 # 低信誉用户更严格 # 综合决策 max_score = max(scores.values()) * weight if max_score > 0.9: return "block", "high_confidence_violation" elif max_score > 0.7: return "replace", "mask_sensitive_content" elif max_score > 0.5: return "flag", "needs_review" else: return "allow", None