Answers are grounded in this paper and cite their sources.
When detecting text generated by Generative Pre-trained Transformer (GPT), the following issues still need to be further addressed. The existing text features and pre-trained language models lack grasp of statistical, local and global semantic information. In response to these shortcomings, this paper quantifies the writing style characteristics of texts, and proposes a Structured Lexical Semantic (SLS) method. This method has contributed to the feature analysis of text generated by artificial intelligence. The SLS uses text statistical features as auxiliary signals to enhance classification ability. Meanwhile, this research utilizes local and global semantic information and proposes a robustly optimized semantic convolutional neural network based on bidirectional encoder representation from Transformers (RoSemCNN). Considering the practical application scenarios of GPT generated text, this paper will focus on detecting text that has been rewritten and polished by GPT. The experimental results on rewriting paragraphs show that the accuracies of SLS and RoSemCNN methods increased by 7.3% and 8.2% compared with the baseline model, verifying their superiority on challenging detection tasks.
No takes yet. Share an insight, caveat, or question.
Jin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: