Turnitin AI生成内容检测原理

# LTI 1.3 集成核心配置示例LTI_CONFIG = {    "issuer": "https://canvas.instructure.com",    "client_id": "turnitin-ai-detect-001",    "platform": "canvas",    "deployment_id": "dep_2024_001",    "target_link_uri": "https://api.turnitin.com/lti/launch",    "scopes": [        "https://purl.imsglobal.org/spec/lti-ags/score.json",        "https://purl.imsglobal.org/spec/lti-nrps/scope/contextmembership.readonly"    ]} # 作业提交后自动触发检测的回调处理def on_submission_received(submission_data):    doc_id = submission_data["document_id"]    # 同时启动查重和AI检测    similarity_task = start_similarity_check(doc_id)    ai_detection_task = start_ai_detection(doc_id)    # 等待两个检测完成后生成融合报告    report = generate_fusion_report(similarity_task, ai_detection_task)    push_report_to_lms(report, submission_data["course_id"])

import mathfrom collections import Counterfrom typing import List, Dict, Tuple # ===== 文本困惑度计算 =====class SimpleLanguageModel:    """基于N-gram的简单语言模型，用于计算困惑度"""    def __init__(self, n: int = 3):        self.n = n        self.ngram_counts: Dict[tuple, int] = {}        self.context_counts: Dict[tuple, int] = {}        self.vocab: set = set()     def train(self, texts: List[str]):        """在人类文本语料上训练N-gram模型"""        for text in texts:            tokens = text.lower().split()            self.vocab.update(tokens)            for i in range(len(tokens) - self.n + 1):                ngram = tuple(tokens[i:i + self.n])                context = tuple(tokens[i:i + self.n - 1])                self.ngram_counts[ngram] = self.ngram_counts.get(ngram, 0) + 1                self.context_counts[context] = self.context_counts.get(context, 0) + 1     def perplexity(self, text: str) -> float:        """计算文本困惑度"""        tokens = text.lower().split()        if len(tokens) < self.n:            return 50.0  # 默认困惑度        log_prob_sum = 0.0        count = 0        for i in range(len(tokens) - self.n + 1):            ngram = tuple(tokens[i:i + self.n])            context = tuple(tokens[i:i + self.n - 1])            ngram_count = self.ngram_counts.get(ngram, 0)            context_count = self.context_counts.get(context, 0)            # 拉普拉斯平滑            prob = (ngram_count + 1) / (context_count + len(self.vocab))            log_prob_sum += math.log(prob)            count += 1        avg_log_prob = log_prob_sum / count        return math.exp(-avg_log_prob) # ===== 突发性计算 =====def burstiness(text: str, model: SimpleLanguageModel) -> float:    """计算文本突发性：各句子困惑度的方差"""    sentences = [s.strip() for s in text.split('.') if s.strip()]    if len(sentences) < 2:        return 0.0    ppls = [model.perplexity(s) for s in sentences]    mean_ppl = sum(ppls) / len(ppls)    variance = sum((p - mean_ppl) ** 2 for p in ppls) / len(ppls)    return variance # ===== AI生成概率分类器 =====def ai_detection_score(text: str, model: SimpleLanguageModel,                       alpha: float = -0.8, beta: float = -0.5,                       gamma: float = 2.0) -> float:    """计算AI生成概率（0-1之间）"""    ppl = model.perplexity(text)    burst = burstiness(text, model)    # 低困惑度和低突发性 -> 更高AI生成概率    z = alpha * ppl + beta * burst + gamma    prob = 1.0 / (1.0 + math.exp(-z))    return prob # ===== 融合报告生成 =====def generate_fusion_report(text: str, model: SimpleLanguageModel) -> Dict:    """生成查重+AI检测融合报告"""    paragraphs = [p.strip() for p in text.split('\n') if p.strip()]    report = {        "total_paragraphs": len(paragraphs),        "similarity_results": [],        "ai_detection_results": [],        "overall_ai_percentage": 0.0    }    ai_para_count = 0    for idx, para in enumerate(paragraphs):        # AI检测（段落级别）        ai_prob = ai_detection_score(para, model)        report["ai_detection_results"].append({            "paragraph_index": idx,            "ai_probability": round(ai_prob, 4),            "flagged": ai_prob > 0.5        })        if ai_prob > 0.5:            ai_para_count += 1        # 查重模拟（实际中调用相似度数据库）        report["similarity_results"].append({            "paragraph_index": idx,            "similarity_score": 0.0,            "matched_sources": []        })    report["overall_ai_percentage"] = round(ai_para_count / max(len(paragraphs), 1) * 100, 1)    return report # ===== 主程序：演示检测流程 =====if __name__ == "__main__":    # 训练数据（模拟人类写作语料）    human_texts = [        "The quick brown fox jumps over the lazy dog near the riverbank.",        "She walked through the forest thinking about her grandmother's stories.",        "Research methodology requires careful consideration of sample size and bias."    ]    model = SimpleLanguageModel(n=3)    model.train(human_texts)     # 待检测文本    test_text = "The quick brown fox jumps over the lazy dog. This is a very predictable sentence."     # 生成融合报告    report = generate_fusion_report(test_text, model)    print(f"总段落数: {report['total_paragraphs']}")    print(f"AI生成比例: {report['overall_ai_percentage']}%")    for r in report["ai_detection_results"]:        flag = "[AI疑似]" if r["flagged"] else "[正常]"        print(f"  段落{r['paragraph_index']}: AI概率={r['ai_probability']} {flag}")