DeepPavlov at SemEval-2024 Task 8: Leveraging Transfer Learning for Detecting Boundaries of Machine-Generated Texts

Anastasia Voznyuk, Vasily KonovalovNAACL 2024

TL;DR

This paper tackles the increasingly important problem of identifying where human-written text transitions to machine-generated text within a single document. The authors developed a novel data augmentation pipeline and fine-tuned DeBERTaV3 models, achieving the best Mean Absolute Error (MAE) score on the SemEval-2024 Task 8, demonstrating the effectiveness of their approach for this nuanced task.

What Problem Does This Paper Address?

The rapid advancement of large language models (LLMs) like GPT-3 and GPT-4 has made it easier than ever to generate coherent and human-like text. While this has numerous benefits, it also raises concerns about misuse, such as generating fake news or devaluing the learning process. Existing methods for detecting machine-generated text often focus on classifying an entire document as either human or AI-written. However, the rise of collaborative human-AI writing necessitates a more granular approach: identifying the precise boundaries where authorship shifts within a text. This is crucial for understanding and potentially mitigating the impact of AI-generated content in various domains. SemEval-2024 Task 8 specifically targets this "boundary detection" problem, challenging researchers to pinpoint these transitions.

Methods

The core of the DeepPavlov team's approach lies in a two-pronged strategy: data augmentation and transfer learning with advanced transformer models.

  • Data Augmentation Pipeline: Recognizing the limited size of the provided training data, the authors devised a pipeline to create more training examples. This involved splitting texts into sentences, then into whitespace-split words, and carefully comparing these to the original text's word splits. The key challenge here was to correctly handle discrepancies, such as missing whitespace or words split by line breaks, to ensure accurate boundary label calculation. This process aimed to generate coherent texts with authorship changes that wouldn't mislead models.
  • Transfer Learning with DeBERTaV3: The researchers fine-tuned several state-of-the-art transformer models, including RoBERTa, Longformer, and DeBERTaV3, on the boundary detection task. DeBERTaV3, with its disentangled attention mechanism for encoding token position and content separately, was particularly promising. The models were trained for token classification, predicting whether each token was part of the human-written prefix or the AI-generated continuation.

Datasets

The primary dataset used for this task was provided by the SemEval-2024 Task 8 organizers. This dataset is derived from the M4 dataset and includes texts from various domains and generators. For training and validation, texts were generated from scientific paper reviews (PeerRead), while the test set also included texts from Outfox, a dataset of LLM-generated student essays, to test robustness against domain shifts. The authors also created an augmented dataset by applying their data augmentation pipeline to the original training data.

Key Insights and Ideas

  • Augmentation is Crucial: The simple yet effective data augmentation pipeline significantly boosted model performance, highlighting the importance of data variety (in text length and boundary positions) for robust boundary detection.
  • DeBERTaV3's Superiority: The DeBERTaV3 models, particularly the large version, consistently outperformed other tested models, demonstrating the advantage of its advanced architecture for this task.
  • Domain Shift Challenges: The error analysis revealed that the model struggled more with texts from the Outfox domain (student essays) compared to the PeerRead domain (scientific reviews). This suggests that differences in formatting, punctuation, and text structuring across domains pose a significant challenge.
  • Anomalies as Hints: Certain anomalies in LLM-generated text, such as excessive repetition, JSON-structured hallucinations, or bizarre formatting (like starting with triple quotes), can actually serve as helpful signals for boundary detection models.

Key Results

The DeepPavlov team achieved the new best Mean Absolute Error (MAE) score on the SemEval-2024 Task 8 leaderboard, with their fine-tuned DeBERTaV3-large model reaching an MAE of 13.38 on the test set using the augmented dataset. This represents an improvement of over 2 points compared to the top-1 submission's MAE of 15.683. The fine-tuned DeBERTaV3-large model achieved an MAE of 15.16 on the original dataset and 13.38 on the augmented dataset, outperforming both RoBERTa and Longformer models across all setups.

Limitations and Future Work

The authors acknowledge that the provided dataset was limited to English texts. Future work could involve training multilingual boundary detection solutions by incorporating data from different languages and using multilingual encoders. Additionally, they note that the test set lacks domain classification labels, making it difficult to evaluate performance on specific domains separately. The error analysis also indicated that the model made most of its mistakes on the Outfox domain, suggesting that further research into domain-specific adaptation or more robust cross-domain generalization techniques is warranted.

Summary updated

BibTeX
@inproceedings{voznyuk2024deeppavlov,
  author = {Anastasia Voznyuk and Vasily Konovalov},
  title = {{DeepPavlov at SemEval-2024 Task 8: Leveraging Transfer Learning for Detecting Boundaries of Machine-Generated Texts}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure