Abstract

This study presents the development of an AI text generation detection tool based on the Transformer model, aiming to enhance the accuracy of AI text generation detection and provide a reference for future research. The tool employs a series of preprocessing steps on the text, including Unicode normalization, conversion to lowercase, removal of non-alphabetic characters using regular expressions, and the addition of spaces around punctuation marks. It also eliminates leading and trailing spaces, replaces consecutive ellipses with a single space, and connects text segments with a specified delimiter. Subsequently, non-alphabetic characters and redundant spaces are removed, and multiple consecutive spaces are replaced with a single space, followed by a second conversion to lowercase.

The deep learning model integrates layers of LSTM, Transformer, and CNN for text classification and sequence labeling tasks. The training and validation sets demonstrate that the model's loss decreased from 0.127 to 0.005, while the accuracy increased from 94.96% to 99.8%, indicating strong detection and classification capabilities for AI-generated text. The confusion matrix and accuracy metrics of the test set reveal a prediction accuracy of 99% for AI-generated text, with a precision of 0.99, recall of 1, and an F1 score of 0.99, achieving a highly accurate classification. The results suggest that this model holds great potential for widespread application in the field of AI text detection.