This paper proposes a sentence-level sequence labeling model to overcome the limitations of AI text detection. While existing document-level classification models struggle to distinguish between mixed or slightly modified AI text, our model utilizes subtle linguistic cues between sentences to detect transitions between AI-generated and human-written text. By combining the latest Transformer model, neural networks (NN), and conditional random fields (CRFs), we achieve precise token-level AI text segmentation. We conduct experiments using two public benchmark datasets, and validate the model's performance through comparisons with existing state-of-the-art models and ablation studies. The source code and processed datasets are available on the GitHub repository.
