TL;DR
This paper explores how Large Language Models (LLMs) with "structure-aware" pre-training can be fine-tuned to distinguish between human-written and machine-generated text. While their models showed modest performance compared to top competitors, they highlight the potential of incorporating linguistic structure for better text identification and suggest future work in more advanced models and hyperparameter tuning.
What Problem Does This Paper Address?
The rapid advancement and accessibility of LLMs like ChatGPT have led to concerns about their potential misuse for generating fake or misleading content. Distinguishing between human and machine-generated text is becoming increasingly challenging, necessitating automated systems to mitigate these risks. SemEval-2024 Task 8 specifically addresses this by providing a benchmark for identifying AI-generated text, including tasks for binary classification (human vs. machine) and multi-class classification (identifying the specific LLM). This paper focuses on the challenge of accurately identifying the origin of text, especially as LLMs become more sophisticated.
Methods
The authors fine-tuned four pre-trained LLMs: ERNIE, SpanBERT, ConvBERT, and XLNet. These models were chosen for their ability to integrate "structure-aware" features during pre-training, meaning they implicitly or explicitly learn about linguistic structures like syntax, semantics, and sentence organization. The fine-tuning process involved standard practices such as data partitioning (train/validation/test), hyperparameter tuning (batch size, learning rate, epochs, weight decay), and using the Hugging Face Transformers library. For classification, the models' output logits were passed through a Softmax function and then an argmax function to determine the final predicted class.
Datasets
The study utilized the dataset provided for SemEval-2024 Task 8, which is an extension of the M4 dataset. This dataset is described as a large-scale, multi-generator, multi-domain, and multi-lingual corpus for machine-generated text detection. It includes texts from various domains (news, code, narratives) and languages (English, Spanish, Chinese, Arabic, Russian, Urdu, Indonesian, Bulgarian). The dataset contains both non-parallel human data and parallel human and machine-generated texts, generated by LLMs such as GPT-4, ChatGPT, GPT3.5, Cohere, and Dolly-v2.
Key Insights and Ideas
- Structure Matters: The core hypothesis is that integrating structural information (lexical, semantic, syntactic) into LLMs can improve their ability to differentiate human and machine writing styles.
- Human Text is Harder to Classify: Confusion matrices reveal that all tested models struggled more with correctly classifying human-written text (high false positive rates) compared to machine-generated text.
- Model Performance Varies: Different LLMs showed varying strengths. For instance, ConvBERT performed well on Subtask A (monolingual), while XLNet showed better performance on Subtask B after re-evaluation.
- Baseline Performance: The baseline model (ROBERTa) achieved a higher accuracy than the authors' best submissions in both Subtask A and Subtask B, indicating the competitive nature of the task.
Key Results
- Subtask A (Monolingual): The authors' best submission achieved an accuracy of 71.5% (using ConvBERT LLM), ranking 87th out of 137 teams. The baseline (ROBERTa) achieved 74% accuracy. Post-competition evaluation with gold labels showed ConvBERT achieving 77.6% accuracy.
- Subtask B: The best submission achieved an accuracy of 38.2% (using ConvBERT LLM), ranking 67th out of 77 teams. The baseline (ROBERTa) achieved 74.6% accuracy. Post-competition evaluation showed XLNet achieving 65.2% accuracy.
- Qualitative Findings: Confusion matrices indicated that while models were generally good at identifying machine-generated text (around 90% accuracy), they struggled with human text classification. For Subtask B, there was significant confusion in classifying human texts, particularly when compared to the "dolly" machine model.
Limitations and Future Work
- Technical Issues: The authors encountered server issues that prevented them from submitting their best-performing models from the validation stages for both subtasks. This is a significant limitation impacting their final ranking.
- Modest Performance: The achieved scores were modest, especially in Subtask B, and below those of the top-ranking teams.
- Future Work: The authors suggest exploring more LLMs that focus on multilevel language and stylistic features. They also propose applying more robust fine-tuning processes with extensive hyperparameter tuning and investigating alternative approaches like Graph Neural Networks for text graph analysis.
Summary updated