Method for detecting generated text based on a maximum mean discrepancy of depth
By constructing a deep composite kernel network and introducing a Wild Bootstrap mechanism, the problems of insufficient capture of high-dimensional semantic differences and sequence correlation in generated text detection are solved, and the stability and generalization ability are improved under various training data conditions, thereby improving the accuracy and robustness of generated text detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-26
AI Technical Summary
Existing generated text detection methods are insufficient in capturing high-dimensional semantic differences, handling sentence sequence correlations within paragraphs, and in terms of stability and generalization ability in scenarios with limited or imbalanced training data, resulting in poor detection performance.
Paragraph-level generated text detection is modeled as a nonparametric two-sample test problem. A deep composite kernel network is constructed to integrate shallow syntactic features and deep semantic features of the text. A Wild Bootstrap mechanism is introduced to generate random weight sequences with autoregressive structures. The deep kernel network is trained by maximizing the unbiased estimation of test power, and the observation statistics are calculated and the detection results are output.
It significantly improves the ability to capture high-dimensional semantic differences in generated text detection, solves the problem of decreased test power caused by sequence correlation, enhances stability under limited or imbalanced training data and generalization ability to unknown generative models, and improves detection accuracy and robustness.
Smart Images
Figure CN122287591A_ABST