TL;DR
This paper introduces a novel architecture for detecting AI-generated text that focuses on syntactic features rather than just semantic content. By aggregating information from all layers of a transformer model and treating text detection as a token-level classification task, the system achieved second place in the multilingual track of SemEval-2024 Task 8, demonstrating strong generalization across unseen domains and languages.
What Problem Does This Paper Address?
The rapid advancement and widespread adoption of large language models (LLMs) like ChatGPT have revolutionized content creation but also introduced significant challenges. The proliferation of AI-generated text raises concerns about misinformation, academic fraud, and the erosion of trust in digital content. Consequently, there's an urgent need for robust methods to reliably distinguish between human-authored and AI-generated text. SemEval-2024 Task 8 specifically targets this challenge with a complex setup involving multiple generators, domains, and languages, evaluated on unseen data, making it a rigorous benchmark for AI text detection systems.
Methods
The authors propose a "syntactically motivated architecture" for binary text classification. Instead of relying solely on the final layer of a transformer model, their approach dynamically aggregates information from all transformer layers. This "layer fusion" technique is inspired by research showing that different layers capture distinct linguistic features, with earlier layers often encoding syntactic information and deeper layers focusing on semantics.
A key innovation is treating text classification as a token-level prediction problem. Each token in the input text is classified as either human-generated or AI-generated. The final classification for the entire sentence is then determined by averaging these token-level predictions. This sequence labeling approach, combined with layer fusion, aims to capture more complex phrasal structures and subtle stylistic differences that might be missed by traditional methods. The model uses an xlm-roberta-large backbone.
Datasets
The experiments were conducted using the dataset provided for SemEval-2024 Task 8. This dataset is described as a large-scale, multi-generator, multi-domain, and multilingual corpus. The machine-generated texts were produced by various LLMs (including ChatGPT, text-davinci-003, Cohere, Dolly-v2, and BLOOMz) across diverse sources and languages (English, Chinese, Urdu, Bulgarian, and Indonesian). The evaluation was performed on an out-of-domain validation set, emphasizing the model's generalization capabilities.
Key Insights and Ideas
- Syntax over Semantics: The paper's preliminary study revealed that semantic embeddings alone are insufficient for distinguishing AI from human text, as generated texts often occupy similar positions in the latent semantic space. This motivated a focus on syntactic features.
- Layer Fusion is Crucial: Aggregating information from all transformer layers, rather than just the final one, significantly boosts performance. This highlights the importance of capturing linguistic nuances at different levels of representation.
- Token-Level Prediction Matters: Treating text detection as a sequence labeling task (classifying each token) proved more effective than traditional sentence-level classification, likely by enabling the model to learn finer-grained distinctions.
- Generalization is Key: The model's strong performance on unseen domains and languages underscores the importance of robust architectures that can generalize beyond their training data, especially in the context of rapidly evolving LLMs.
Key Results
The proposed system achieved an accuracy of 95.8% on the SemEval-2024 Task 8 test leaderboard for the multilingual track (Subtask A). This secured them the 2nd position out of 69 participants in that track, narrowly missing the top spot by just 0.1%. Notably, their approach used a smaller-scale transformer model compared to many other top-performing systems that relied on fine-tuned LLMs. Ablation studies showed that omitting sequence labeling resulted in a 6-point decrease in accuracy, while excluding layer fusion led to a more substantial drop of over 10 points in F1 score.
Limitations and Future Work
The paper acknowledges that while their model achieved strong performance, it still relies on a pre-trained transformer backbone. Future work could explore even more efficient architectures or investigate how to further improve generalization to entirely novel LLMs or generation techniques. The authors also note that their system uses a smaller model compared to some competitors, implying that further scaling or different architectural choices might yield even better results, though at a higher computational cost. The learned weight analysis suggests that specific layers (3-6 and 20-22) play a pivotal role, which could inform future architectural designs.
Summary updated