Research on ChatGPT generated text detection model based on phonetic feature extraction and semantic features.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• The proliferation of large language models (LLMs) like ChatGPT presents significant challenges in distinguishing AI-generated text from human writing.
• Current detection methods, primarily based on semantic analysis, lack robustness against paraphrased or rewritten AI content.

Purpose of the Study:

• To develop an integrated framework for robust AI-generated text detection.
• To enhance detection capabilities by combining deep contextual semantics with auxiliary surface-level features.

Main Methods:

• Utilized a RoBERTa encoder for deep contextual semantic embeddings.
• Incorporated a convolutional neural network (CNN) for multi-scale representation aggregation.
• Integrated surface-level features: pronunciation cues, structural, lexical, and readability descriptors.

Main Results:

• The proposed RoBERTa-CNN framework significantly outperformed existing baselines on the HAGTC and ChatGPT abstract datasets.
• Achieved superior accuracy and F1 scores in detecting AI-generated text, including rewritten content.
• Ablation studies confirmed the performance gains from fusing multiple feature types.

Conclusions:

• Combining contextual semantic embeddings with auxiliary surface features offers a practical and effective approach for AI-generated text detection.
• The integrated framework demonstrates enhanced robustness, particularly for detecting paraphrased or rewritten AI content.
• Feature-level fusion is a key strategy for advancing the accuracy and reliability of AI text detection systems.