Fine-Grained Machine-Generated Text Detection
Keywords: Machine-Generated Text Detection, Fine-grained Classification, Mixture of Experts
Abstract: Machine-Generated Text (MGT) detection identifies whether a given text is human-written or machine-generated. However, this can result in detectors that would flag paraphrased or translated text as machine-generated. Fine-grained classification that separates the different types of machine text is valuable in real-world applications, as different types of MGT convey distinct implications. For example, machine-generated articles are more likely to contain misinformation, whereas paraphrased and translated texts may improve understanding of human-written text. Despite this benefit, existing studies consider this a binary classification task, either overlooking machine-paraphrased and machine-translated text entirely or simply grouping all machine-processed text into one category. To address this shortcoming, this paper provides an in-depth study of fine-grained MGT detection, categorizing input text into four classes: human-written, machine-generated, machine-paraphrased, and machine-translated. A key challenge is the performance drop on out-of-domain texts due to the variability in text generators, especially for translated or paraphrased text. We introduce a RoBERTa-based Mixture of Detectors (RoBERTa-MoD), which leverages multiple domain-optimized detectors for more robust and generalized performance. We offer theoretical proof that our method outperforms a single detector, and experimental findings demonstrate a 5--9\% improvement in mean Average Precision (mAP) over prior work on six diverse datasets: GoodNews, VisualNews, WikiText, Essay, WP, and Reuters. Our code and data will be publicly released upon acceptance.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8264
Loading