Team MLab at SemEval-2024 Task 8: Analyzing Encoder Embeddings for Detecting LLM-generated Text

Kevin Li, Kenan Hasanaliyev, Sally Zhu, George Altshuler, Alden Eberts, Eric Chen, Kate Wang, Emily Xia, Eli Browne, Ian ChenNAACL 2024

TL;DR

This paper tackles the growing challenge of distinguishing human-written text from machine-generated text, especially when identifying the specific Large Language Model (LLM) responsible. The authors found that their DeBERTa model significantly outperformed a RoBERTa baseline in this task, demonstrating the potential of advanced encoder models for AI text detection. Visualizations further revealed DeBERTa's ability to group LLM outputs, offering insights into how these models "reason" about text.

What Problem Does This Paper Address?

The explosive popularity of LLMs like ChatGPT has brought immense benefits but also significant challenges, particularly in distinguishing AI-generated content from human writing. This is crucial for combating misinformation, ensuring accountability, and understanding the societal impact of LLMs, which are increasingly integrated into various professions. The SemEval-2024 Task 8 specifically aimed to develop AI systems that can not only detect machine-generated text but also identify the specific LLM that produced it. This paper focuses on Subtask B of this task, which involves this more granular identification.

Methods

The core of this work involves comparing two powerful transformer-based encoder models: RoBERTa and DeBERTa.

  • RoBERTa: Used as the baseline, this model is an optimized version of BERT, known for its improved pre-training techniques.
  • DeBERTa: This model, an upgrade to BERT and RoBERTa, incorporates disentangled attention and an enhanced mask decoder, along with adversarial training during fine-tuning. The authors hypothesized that DeBERTa would outperform RoBERTa and proceeded to fine-tune it on the task's dataset.
  • Hyperparameter Tuning: For DeBERTa, the authors systematically explored various learning rates, batch sizes, epoch counts, and weight decay constants to find the optimal configuration.
  • Model Interpretability: To understand DeBERTa's decision-making process, the authors employed dimensionality reduction techniques (PCA and t-SNE) on the model's pooled outputs. These visualizations aimed to reveal how the model clusters different types of text.

Datasets

The paper primarily discusses its participation in SemEval-2024 Task 8, which provided the datasets for training and evaluation. Subtask B specifically involved identifying text generated by various LLMs, including ChatGPT, Cohere, Davinci, Bloomz, and Dolly, alongside human-written text. The authors mention using 18,000 sentences from the test data for their interpretability analysis.

Key Insights and Ideas

  • DeBERTa's Superiority: The DeBERTa model demonstrated a significant advantage over the RoBERTa baseline in identifying LLM-generated text and, crucially, in distinguishing between different LLMs. This highlights the evolving capabilities of more advanced encoder architectures.
  • LLM "Mannerisms": The paper suggests that certain LLMs, like Bloomz (trained multilingually), might exhibit more discernible "mannerisms" in their text generation, making them easier to detect. This could be linked to the model's training data and multilingual capabilities.
  • Davinci's Anomaly: The Davinci model presented an interesting case, performing well on validation sets but poorly on the test set for DeBERTa. This discrepancy might be due to Davinci being an earlier model, potentially influencing newer models or having characteristics that are harder for newer detection models to pinpoint.
  • Visualizing AI "Thought": The PCA and t-SNE visualizations provided compelling evidence that DeBERTa can effectively cluster outputs from different LLMs and human text. This offers a glimpse into the internal representations learned by these models, aiding in understanding their detection mechanisms.

Key Results

  • Performance Boost: The DeBERTa model achieved a final accuracy of 0.827 on Subtask B, significantly outperforming the RoBERTa baseline by over 10% in leaderboard accuracy.
  • High Metrics: The DeBERTa model achieved weighted average scores of 0.98633 precision, 0.98599 recall, and 0.98599 F1-score on the validation set.
  • RoBERTa Performance: The RoBERTa baseline achieved weighted scores of 0.97979 precision, 0.97909 recall, and 0.97909 F1-score, which, while high, were notably lower than DeBERTa's.
  • Confusion Matrix Insights: The confusion matrix revealed specific misclassification patterns, such as a significant number of Cohere predictions being misattributed to Davinci.
  • Visualizations: Figures 1 and 2 show distinct clusters for human text and various LLM outputs, with Davinci text appearing more scattered across different clusters, indicating potential challenges in its detection.

Limitations and Future Work

  • Davinci's Performance: The paper acknowledges the poor performance of Davinci on the DeBERTa test set and speculates it might be due to Davinci being an older model, potentially making it harder for newer models to detect.
  • ChatGPT's Recall: ChatGPT consistently ranked low in precision and F1-score but had high recall, suggesting it was easily detected as AI-generated but often misclassified as another AI source. The authors suggest swapping ChatGPT for GPT-4 could yield different results, pointing to a future direction.
  • Visualization Parameter Sensitivity: The authors note that the visualization parameters might influence the clarity of the clusters, suggesting that different configurations could yield different insights.
  • Future Work: The authors suggest exploring GPT-4 for improved detection and further investigating the influence of training data and timelines on model detectability. Visualizing AI thought processes is also highlighted as a valuable avenue for improving AI systems.

Summary updated

BibTeX
@inproceedings{li2024team,
  author = {Kevin Li and Kenan Hasanaliyev and Sally Zhu and George Altshuler and Alden Eberts and Eric Chen and Kate Wang and Emily Xia and Eli Browne and Ian Chen},
  title = {{Team MLab at SemEval-2024 Task 8: Analyzing Encoder Embeddings for Detecting LLM-generated Text}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure