M4: Multi-generator, Multi-domain, and Multi-lingual Black-Box Machine-Generated Text Detection

Large language models (LLMs) are becoming mainstream and easily accessible, ushering in an explosion of machine-generated content over various channels, such as news, social media, question-answering forums, educational, and even academic contexts. Recent LLMs, such as ChatGPT and GPT-4, generate remarkably fluent responses to a wide variety of user queries. The articulate nature of such generated text makes LLMs attractive for replacing human labor in many scenarios. However, this has also resulted in concerns regarding their potential misuse, such as spreading misinformation and causing disruptions in the education system. Since humans perform only slightly better than chance when classifying machine-generated vs. human-written text, there is a need to develop automatic systems to identify machine-generated text with the goal of mitigating its potential misuse.

We construct the dataset M4, and additionally study the performance of automatic detectors from various perspectives: (1) different detectors across different domains for a specific generator; (2) different detectors across different generators for a specific domain; (3) interactions of domains and generators in a multilingual setting, and (4) the performance of the detector on data generated from different time domains. From this, we draw some interesting observations.

• Data
• Models
• Evaluation
• Citation

Data

Here are the statistics information about the M4 dataset. It will be further extended in SemEval 2014 shared task 8 with surprising generators, domains and languages.

Models

• RoBERTa, ELECTRA and XLM-R detector: fine-tune pre-trained RoBERTa, ELECTRA and XLM-R to detect machine-generated text.
• Logistic Regression with GLTR Features
• SVM classifier with Stylistic/NEws LAndscape (NELA) Features

Evaluation

We evaluate detectors in five settings:

Same-Generator, Cross-Domain

Accuracy of cross-domain experiments: given generations from ChatGPT (top) or davinci (bottom),train on a single domain and test across domains across five detectors.

Same-Domain, Cross-Generator

Accuracy of cross-generator experiments: train and test on arXiv (top) and Wikipedia (bottom) across five detectors, over single machine-text generator vs human.

GPTZero

Zero-shot detection with GPTZero: recall (Rec) and F1-score with respect to generators and domains.

Multilingual Evaluation

Accuracy (%) based on XLM-R on test sets across different languages over ChatGPT and davinci-003.

Time Domain Evaluation

Impact of ChatGPT updating by time. Accuracy (Acc), Precision (Prec), Recall and F1 scores(%) with respect to machine generations for Reddit-ELI5 from HC3 and M4 datasets based on XLM-R. HC3 dataset prompt ChatGPT earlier than M4 (Jan vs May).

Citation

Please cite us if you use our data or models, see EACL'2024 paper -- Best Resource Paper Award:

@inproceedings{wang-etal-2024-m4,
    title = "M4: Multi-generator, Multi-domain, and Multi-lingual Black-Box Machine-Generated Text Detection",
    author = "Wang, Yuxia  and
      Mansurov, Jonibek  and
      Ivanov, Petar  and
      Su, Jinyan  and
      Shelmanov, Artem  and
      Tsvigun, Akim  and
      Whitehouse, Chenxi  and
      Mohammed Afzal, Osama  and
      Mahmoud, Tarek  and
      Sasaki, Toru  and
      Arnold, Thomas  and
      Aji, Alham  and
      Habash, Nizar  and
      Gurevych, Iryna  and
      Nakov, Preslav",
    editor = "Graham, Yvette  and
      Purver, Matthew",
    booktitle = "Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2024",
    address = "St. Julian{'}s, Malta",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.eacl-long.83",
    pages = "1369--1407",
    abstract = "Large language models (LLMs) have demonstrated remarkable capability to generate fluent responses to a wide variety of user queries. However, this has also raised concerns about the potential misuse of such texts in journalism, education, and academia. In this study, we strive to create automated systems that can detect machine-generated texts and pinpoint potential misuse. We first introduce a large-scale benchmark M4, which is a multi-generator, multi-domain, and multi-lingual corpus for machine-generated text detection. Through an extensive empirical study of this dataset, we show that it is challenging for detectors to generalize well on instances from unseen domains or LLMs. In such cases, detectors tend to misclassify machine-generated text as human-written. These results show that the problem is far from solved and that there is a lot of room for improvement. We believe that our dataset will enable future research towards more robust approaches to this pressing societal problem. The dataset is available at https://github.com/mbzuai-nlp/M4",
}

Shared task using an extension of this dataset

We also have a shared using this data: SemEval-2024 task 8