Machine-generated text authorship attribution detection method based on multi-source feature fusion

By employing a multi-source feature fusion method and utilizing pre-trained language models and techniques such as sparse routing, TF-IDF weights, and singular value decomposition, the problem of insufficient generalization ability and inadequate feature representation in machine-generated text author attribution detection in cross-domain scenarios is solved, achieving more stable author attribution detection.

CN122221038BActive Publication Date: 2026-07-24DALIAN UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN Β· China
Patent Type
Patents(China)
Current Assignee / Owner
DALIAN UNIV OF TECH
Filing Date
2026-05-19
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing methods for detecting author attribution in machine-generated text lack generalization ability in cross-domain scenarios, fail to fully express single features, have coarse fusion mechanisms, and struggle to stably characterize the differences in surface style levels such as function word combinations and local structural rhythm among different generation models.

Method used

A multi-source feature fusion method is adopted. High-dimensional hidden layer representation sequences are extracted through the pre-trained language model RoBERTa. By combining private domain expert networks and shared expert networks, style representations are constructed using sparse routing, TF-IDF weights and singular value decomposition. Semantic and style features are fused, and a residual fusion method is used to generate a comprehensive discriminative representation.

Benefits of technology

It achieves more stable and topic-biased author attribution detection for machine-generated text in cross-domain scenarios, improving detection accuracy and robustness, and avoiding problems such as insufficient generalization ability and inadequate expression of single features.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122221038B_ABST
    Figure CN122221038B_ABST
Patent Text Reader

Abstract

The application discloses a machine-generated text author attribution detection method based on multi-source feature fusion, which combines a shared expert network and a private domain expert network to extract and fuse semantic features of machine-generated text, so that accurate feature extraction is realized; an alignment style representation is constructed by using a constructed sub-linear TF-IDF weight structure, and is mapped into a low-dimensional dense vector through truncated singular value decomposition, so that the semantic representation is taken as a trunk, and the style representation is taken as compensation; the introduction intensity of the style feature is dynamically adjusted through a gating factor, high-order semantic discrimination ability and surface style robustness are considered, so that a more stable and more anti-topic shift machine-generated text author is obtained, and problems such as insufficient generalization ability, insufficient single feature expression and rough fusion mechanism are avoided.
Need to check novelty before this filing date? Find Prior Art