Scholarlink

<p>With the emergence of ChatGPT, an advanced generative AI tool, maintaining academic integrity has become a significant challenge for educators. This paper presents strategies to tackle this issue, focusing on the RoBERTa-CNN model for continuous value predictions, particularly for detecting AI-generated text. The model utilizes the RoBERTa transformer architecture to generate contextualized token embeddings, which are aggregated using mean pooling to create sequence-level representations. The study provides an in-depth mathematical framework for the hybrid RoBERTa-CNN model, covering key components such as input tokenization, self-attention mechanisms, pooling strategies, and classification transformations. The performance outperforms state-of-the-art results, with the RoBERTa-CNN model showing a 1-25% improvement over BERT. The model achieves 100% overall accuracy in identifying human-written text, significantly reducing false positives and ensuring that genuine human content is not misclassified as AI-generated. Evaluation metrics validate its effectiveness in minimizing false negatives, including Recall, F2 score, perplexity, human evaluation, BLEU, ROUGE, and diversity scores. The proposed hybrid transformer-based language model and Machine Learning (ML) classifier enable high-accuracy detection of AI-generated text, ensuring that human-written text are correctly identified. Comparative analysis with trending AI text detection tools, such as DetectGPT, GPTZero, Copyleaks, Fast-DetectGPT, and OpenAI’s models, demonstrates superior performance. A novel n-gram bag-of-words (BOW) discrepancy language model is introduced, providing educators with valuable tools to uphold academic integrity in the age of advanced AI technologies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detecting AI-generated text in high-resource languages: developing a RoBERTa-CNN hybrid model for academic integrity challenge

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Manish Prajapati,
  • Santos Kumar Baliarsingh,
  • Prabhu Prasad Dev

摘要

With the emergence of ChatGPT, an advanced generative AI tool, maintaining academic integrity has become a significant challenge for educators. This paper presents strategies to tackle this issue, focusing on the RoBERTa-CNN model for continuous value predictions, particularly for detecting AI-generated text. The model utilizes the RoBERTa transformer architecture to generate contextualized token embeddings, which are aggregated using mean pooling to create sequence-level representations. The study provides an in-depth mathematical framework for the hybrid RoBERTa-CNN model, covering key components such as input tokenization, self-attention mechanisms, pooling strategies, and classification transformations. The performance outperforms state-of-the-art results, with the RoBERTa-CNN model showing a 1-25% improvement over BERT. The model achieves 100% overall accuracy in identifying human-written text, significantly reducing false positives and ensuring that genuine human content is not misclassified as AI-generated. Evaluation metrics validate its effectiveness in minimizing false negatives, including Recall, F2 score, perplexity, human evaluation, BLEU, ROUGE, and diversity scores. The proposed hybrid transformer-based language model and Machine Learning (ML) classifier enable high-accuracy detection of AI-generated text, ensuring that human-written text are correctly identified. Comparative analysis with trending AI text detection tools, such as DetectGPT, GPTZero, Copyleaks, Fast-DetectGPT, and OpenAI’s models, demonstrates superior performance. A novel n-gram bag-of-words (BOW) discrepancy language model is introduced, providing educators with valuable tools to uphold academic integrity in the age of advanced AI technologies.