TrustAI at SemEval-2024 Task 8: A Comprehensive Analysis of Multi-domain Machine Generated Text Detection Techniques

Ashok Urlana, Aditya Saibewar, Bala Mallikarjunarao Garlapati, Charaka Vinayak Kumar, Ajeet Singh, Srinivasa Rao ChalamalaNAACL 2024

TL;DR

This paper presents a comprehensive analysis of various techniques for detecting machine-generated text across diverse domains and languages for SemEval-2024 Task 8. The authors highlight that ensemble methods excel in mono-lingual detection, while models fine-tuned on GPT2-generated text perform best in multi-class classification, achieving up to 86.9% accuracy.

What Problem Does This Paper Address?

The rapid advancement of Large Language Models (LLMs) has led to an explosion in their ability to generate fluent and coherent text. While this offers numerous benefits, it also raises significant concerns regarding the spread of misinformation, fake news, and the potential for personal information leakage. Detecting machine-generated text is crucial to mitigate these risks. However, LLM-generated text often closely mimics human writing in terms of grammar, coherence, and factual knowledge, making detection a challenging task. This paper addresses the need for robust and versatile machine-generated text detection methods, particularly across multi-domain and multi-lingual contexts, as required by the SemEval-2024 Task 8.

Methods

The authors explore three main categories of techniques for machine-generated text detection:

  • Statistical Methods: This includes traditional machine learning models like Logistic Regression (LR), Support Vector Machines (SVM), Naive Bayes, LightGBM, and MLP. They also experimented with ensemble approaches combining these models.
  • Neural Methods: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks were employed, utilizing FastText embeddings for hierarchical pattern capture.
  • Pre-trained Models: A range of pre-trained transformer-based language models were fine-tuned, including BERT, RoBERTa, DistilRoBERTa, and XLM-RoBERTa.

The study also details two pre-processing strategies: one involving heuristic-based pre-processing and sub-word removal (Version-1), and another solely using heuristic-based pre-processing (Version-2).

Datasets

The study utilizes the dataset provided for SemEval-2024 Task 8, which is an extension of the M4 dataset. This dataset includes both machine-generated text (from various LLMs like ChatGPT, BLOOMz, text-davinci-003, etc.) and human-written text. The human-written text is sourced from diverse platforms such as Wikipedia, WikiHow, arXiv, PeerRead, Reddit, Baike, Web questions, and news articles in multiple languages. The dataset is structured for three subtasks:

  • Subtask A (Mono-lingual): Binary classification of English text as human-written or machine-generated.
  • Subtask A (Multi-lingual): Binary classification of text in English, Chinese, Russian, Urdu, Indonesian, Arabic, and Bulgarian.
  • Subtask B: Multi-class classification to distinguish text from specific sources: 'human', 'chatGPT', 'cohere', 'davinci', 'bloomz', and 'dolly'.

Key Insights and Ideas

  • Ensemble Power for Mono-lingual: Ensemble models, particularly those combining Naive Bayes, SGDClassifier, and LightGBM, proved highly effective for mono-lingual machine-generated text detection.
  • GPT2-tuned Models for Multi-class: For the multi-class classification task (Subtask B), models fine-tuned on GPT2-generated text demonstrated superior performance.
  • Data Pre-processing Matters: The choice of pre-processing (Version-1 vs. Version-2) had a noticeable impact on the performance of pre-trained models, suggesting that how data is cleaned and prepared is critical.
  • Generalization Challenges: A notable discrepancy was observed between development set performance and test set performance for some pre-trained models. This suggests potential overfitting to the training data, possibly due to differing data sources between training and testing splits.
  • Multi-lingual Data Nuances: The multi-lingual dataset presented challenges, including the presence of empty samples, which required careful handling during model training.

Key Results

  • Subtask A (Mono-lingual): The authors achieved an accuracy of 86.9% on the test set using an ensemble model comprising Naive Bayes, SGDClassifier, and LightGBM.
  • Subtask A (Multi-lingual): Fine-tuning BERT Multilingual Base with Version-2 pre-processed data yielded 73.8% accuracy on the test set.
  • Subtask B: The RoBERTa Base OpenAI Detector achieved 83.7% accuracy on the test set, while DistilRoBERTa Base achieved 79.1% accuracy and secured 17th rank.
  • Statistical models generally performed modestly on the development set but generalized better to the test set compared to some pre-trained models.

Limitations and Future Work

The authors acknowledge several limitations:

  • Computational Constraints: Experiments were limited by computational resources, preventing the evaluation of very large language models.
  • Generalization Gap: Some experimental methods showed a significant drop in performance from the development set to the test set, indicating a potential lack of generalization. This is attributed to potential overfitting on training data due to differing data sources.

Future work could involve exploring larger language models, investigating more advanced pre-processing techniques, and developing methods to improve the generalization capabilities of models across diverse data distributions. Further analysis into the specific characteristics of machine-generated text that lead to detection failures would also be beneficial.

Summary updated

BibTeX
@inproceedings{urlana2024trustai,
  author = {Ashok Urlana and Aditya Saibewar and Bala Mallikarjunarao Garlapati and Charaka Vinayak Kumar and Ajeet Singh and Srinivasa Rao Chalamala},
  title = {{TrustAI at SemEval-2024 Task 8: A Comprehensive Analysis of Multi-domain Machine Generated Text Detection Techniques}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure