TL;DR
This paper introduces a multi-task learning architecture for detecting AI-generated text fragments in scientific papers, achieving a 9% improvement over baseline methods. By jointly training binary and multi-class classifiers, the system effectively identifies subtle AI-generated content, reaching a 0.96 F1-score on a challenging competition dataset.
What Problem Does This Paper Address?
The rapid advancement of AI text generation models, like ChatGPT and Llama, has led to the creation of highly fluent and human-like text. While beneficial for many applications, this poses a significant challenge for scientific integrity, as it becomes increasingly difficult to distinguish between human-written and AI-generated content. This can lead to issues like plagiarism, fake studies, and misinformation within academic literature. The DAGPap24 competition specifically aimed to address this by tasking participants with identifying AI-generated token-level fragments within scientific documents, a nuanced problem that requires precise boundary detection.
Methods
The authors propose a multi-task learning architecture built upon a BERT-like encoder. The core innovation lies in the use of two distinct "heads" that process the token embeddings from the encoder:
- Binary Classifier (Linear A): This head is trained to distinguish between human-written and any form of AI-generated text.
- Multi-class Classifier (Linear B): This head further categorizes AI-generated text into one of four predefined classes (ChatGPT, NLTK-synonym replacement, summarized model, or a general AI-generated category).
These two heads are trained jointly, with the loss function aggregating errors from both. This multi-task approach is intended to improve generalization and robustness. For inference, the system first uses the binary classifier. If the confidence in AI-generated content exceeds a certain threshold, the multi-class classifier is then employed. The authors also experimented with varying input sequence lengths and different BERT-like encoders (SciBERT, XLNet, Mistral with QLoRA) to optimize performance.
Datasets
The primary dataset used in this work is the DAGPap24 competition dataset. This dataset consists of scientific texts where specific intervals have been replaced by AI-generated content. The dataset includes annotations, tokenizations, and labels for each token, categorizing them into human-written or one of three AI-generated types. The paper mentions that the dataset was balanced to have an equal number of human-written texts and artificially generated samples.
Key Insights and Ideas
- Multi-task learning for robustness: Jointly training a binary and a multi-class classifier for AI text detection proved more effective than a single classifier, leading to improved generalization and reduced overfitting.
- Context is crucial for subtle detection: AI-generated text can appear mid-sentence, highlighting the importance of considering token context for accurate classification.
- Sequence length matters: The optimal input sequence length for the BERT-like encoder was found to be 350 token-words, balancing information retention with model constraints.
- Domain-specific encoders are beneficial: SciBERT, a BERT model pre-trained on scientific text, served as a strong baseline and ultimately the chosen encoder due to its effectiveness in this domain.
Key Results
The proposed multi-task architecture achieved significant improvements:
- Development Set: A 9% improvement in average macro F1-score was observed, increasing from a baseline of 0.86 to 0.95.
- Closed Test Set: The system achieved a score of 0.96 on the competition's closed test set, securing 5th place.
- Optimal Settings: The best performance was achieved using SciBERT as the encoder, an input sequence length of 350 token-words, and a dropout value of 0.7. A threshold of 0.55 for the binary classifier also contributed to improved results.
Limitations and Future Work
- Limitations: The paper acknowledges the constraint of input sequence length for BERT-like models, which can lead to information loss if not managed properly. The authors also note that the dataset, while challenging, might not encompass all possible AI generation methods.
- Future Work: The authors suggest exploring class balancing by adding weights to the loss function during pre-training. They also express interest in applying this approach to larger datasets with classes defined by more common text generation methods. Analyzing the distribution of metrics across different classes is also a future direction.
Summary updated