Papilusion at DAGPap24: Paper or Illusion? Detecting AI-generated Scientific Papers

Nikita Andreev, Alexander Shirnin, Vladislav Mikhailov, Ekaterina ArtemovaACL 2024

TL;DR

This paper introduces "Papilusion," an AI-generated scientific text detector that achieved a top-tier performance on the DAGPap24 challenge. By employing an ensemble of fine-tuned DeBERTa models and addressing a critical tokenization issue post-competition, they significantly boosted their F1 score, demonstrating the effectiveness of their approach for identifying AI-generated scientific content.

What Problem Does This Paper Address?

The proliferation of AI language models has raised concerns about the potential for fabricating scientific publications and undermining the integrity of the peer-review process. While detecting AI-generated text is an active research area, the DAGPap24 challenge specifically targets a more nuanced task: classifying individual tokens within scientific texts as either human-written or AI-generated. This token-level detection is crucial for understanding how AI might be used as a writing assistant, subtly altering human-written content rather than producing entirely new pieces. The paper aims to develop a robust system to tackle this challenge, contributing to the broader effort of maintaining scientific authenticity.

Methods

Papilusion employs an ensemble-based approach, a common strategy for improving model robustness and performance. The core of their method involves fine-tuning multiple encoder-only sequence tagger models, specifically variations of the DeBERTaV3 architecture. These models are trained independently with different hyperparameters and configurations (e.g., varying frozen layers, input sequence lengths) to create a diverse ensemble.

The key innovation lies in their post-competition analysis. They discovered a tokenization issue in their initial implementation where sequences were chunked by word count rather than token count, leading to information loss. Correcting this, along with further hyperparameter tuning and ensemble optimization, led to a substantial performance improvement. The final ensemble uses a majority voting mechanism to aggregate predictions from the best-performing individual models.

Datasets

The primary dataset used is the DAGPap24 dataset. This dataset consists of scientific texts that are a mix of human-written, modified through synonym replacement, produced by ChatGPT, and summarized. The dataset is divided into training, development, and test sets, with specific sizes for each. The paper does not introduce a new benchmark but rather addresses a task within the existing DAGPap24 framework.

Key Insights and Ideas

  • Token-level detection is a more realistic scenario: The paper highlights that AI is increasingly used as a writing assistant, making token-level detection more relevant than simply classifying entire documents.
  • Ensembling is powerful, but careful implementation matters: While ensembling multiple models is a standard technique, the authors found that fixing a subtle tokenization bug had a more significant impact than many hyperparameter optimizations. This underscores the importance of meticulous data processing.
  • DeBERTa models are highly effective, even smaller ones: Even the smallest DeBERTa-Xsmall model achieved an impressive F1 score above 98, suggesting that the task might not require extremely large models, especially with proper fine-tuning and data handling.
  • Synonym replacement can simplify the task: The authors hypothesize that synonym replacement, while intended to make detection harder, might actually create more distinct patterns that AI models can learn, potentially simplifying the classification task for certain classes.

Key Results

  • Competition Performance: Papilusion ranked 6th in the DAGPap24 competition with an F1 score of 89.83.
  • Post-Competition Improvement: After fixing the tokenization issue and further optimizing their models, their F1 score on the official test set reached 99.46, an improvement of +9.63.
  • Single Model Performance: The best single DeBERTa-large model achieved a score of 99.21 with 18 out of 24 layers frozen.
  • Ensemble Performance: The majority voting ensemble of their best DeBERTa-large models achieved 99.46, placing them second overall based on competition results.

Limitations and Future Work

  • Dataset Characteristics: The authors suggest that the DAGPap24 dataset might be less complex than initially anticipated due to its origin and the nature of synonym replacement, which could have made the task easier for AI models. They hypothesize that the construction process might have relied on basic generation rules and lacked sufficient filtering of simple samples.
  • Future Work: While not explicitly detailed, the paper's findings suggest that further exploration into more sophisticated data augmentation techniques, robust tokenization strategies, and potentially exploring different model architectures could lead to even more advanced AI-generated text detectors. The authors also implicitly point to the ongoing challenge of creating truly robust detectors that can generalize across various AI generation methods and subtle modifications.

Summary updated

BibTeX
@inproceedings{andreev2024papilusion,
  author = {Nikita Andreev and Alexander Shirnin and Vladislav Mikhailov and Ekaterina Artemova},
  title = {{Papilusion at DAGPap24: Paper or Illusion? Detecting AI-generated Scientific Papers}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure