TL;DR
This paper explores using text perplexity as a feature to distinguish human-written text from machine-generated text. The authors found that while perplexity alone isn't perfect, combining it with word embeddings from a fine-tuned XLM-RoBERTa-Large model shows promise, though performance on the test set was unexpectedly lower than anticipated.
What Problem Does This Paper Address?
The rapid advancement of generative AI models means that machine-generated text is becoming increasingly indistinguishable from human writing. This poses significant challenges, from the spread of misinformation to privacy concerns. The paper addresses the critical need for robust systems that can reliably identify whether a piece of text was authored by a human or an AI. This is crucial for maintaining trust and integrity in digital communication.
Methods
The SINAI team developed three systems for the SemEval-2024 Task 8, which focuses on machine-generated text detection.
- System One (Fine-tuning): This system solely relies on fine-tuning the XLM-RoBERTa-Large language model on the training data. This is a standard approach in NLP for adapting pre-trained models to specific tasks.
- System Two (Perplexity): This system uses text perplexity as the sole classification signal. Perplexity, a measure of how well a probability model predicts a sample, is hypothesized to be lower for machine-generated text than for human-written text. A threshold was established to classify texts.
- Proposed System (Fine-tuning + Perplexity): This is the core contribution. It combines the word embeddings generated by the fine-tuned XLM-RoBERTa-Large model (from System One) with the perplexity score of the input text (from System Two). A multimodal toolkit was used to fuse these two feature sets before the final classification layer.
Datasets
The paper utilizes datasets for the SemEval-2024 Task 8, which includes both monolingual (English) and multilingual classification tasks.
- Monolingual English Task:
- Training: 119,757 instances
- Evaluation: 5,000 instances
- Multilingual Task:
- Training: 172,417 instances
- Evaluation: 4,000 instances
- The multilingual dataset comprises primarily English text (77.48%), with Bulgarian as a secondary language, and also includes Chinese, Indonesian, and Urdu. The evaluation dataset includes Russian, German, and Arabic.
Each instance includes the text and its source (e.g., Wikihow, Wikipedia, Reddit, Arxiv, Peerread). A gold label indicates whether the text is machine-generated (1) or human-written (0). The datasets are noted to have an even distribution of human and machine-generated texts in the training and development corpora.
Key Insights and Ideas
- Perplexity as a Signal: The central hypothesis is that perplexity can serve as a valuable feature for detecting machine-generated text. The intuition is that human language exhibits a different probability distribution than AI-generated text.
- Human vs. Machine Perplexity: The study found a notable difference in perplexity scores between human-written and machine-generated texts, supporting the hypothesis. Machine-generated texts generally had lower perplexity.
- Fusion of Features: Combining word embeddings (capturing semantic and syntactic information) with perplexity (capturing linguistic predictability) proved to be a promising strategy, outperforming individual approaches during development.
- Model Discrepancy: A significant finding was the unexpected difference in perplexity scores between the training/development sets and the test set. This suggests that the large language models used to generate the test data might be more sophisticated or different from those used for the training data, impacting the effectiveness of the perplexity feature.
Key Results
- Development Phase:
- System One (fine-tuning) achieved an accuracy of 0.8002 on the development set, outperforming the task baseline.
- System Two (perplexity) achieved an accuracy of 0.6894, which was lower than the baseline but still suggested perplexity's potential as a discriminatory feature.
- The Proposed System (fusion) achieved strong results, with 0.8698 for monolingual and 0.6789 for multilingual on the development set, significantly outperforming the task baseline in the monolingual scenario.
- Final Test Phase:
- The final results on the test set were unexpectedly lower than the development phase. The monolingual task accuracy was 0.744631, and the multilingual task accuracy was 0.801689.
- Both final results were lower than the competitor baselines (0.8846 for monolingual test and 0.8088 for multilingual test).
- The paper notes that the proposed system had no false negatives on the test set, meaning it correctly identified all machine-generated texts. However, there were false positives (human texts misclassified as machine-generated).
Limitations and Future Work
- Perplexity Metric Dependency: The authors acknowledge that the perplexity metric is dependent on the reference large language model used for its calculation. The disparity between models used for training and testing data is identified as a key limitation.
- Model Sophistication: The unexpected performance drop on the test set is attributed to the potential use of more sophisticated or different LLMs for test data generation.
- Future Work: The authors suggest exploring the influence of different reference large language models for perplexity calculation. They also propose investigating the possibility of combining perplexity scores calculated using a diverse set of LLMs to create a more robust feature. Further work is also needed to reduce false positives.
Summary updated