TL;DR
This paper introduces AIpom, a novel two-stage pipeline system for detecting boundaries between human-written and AI-generated text. By combining a fine-tuned decoder-only language model with encoder-only sequence taggers, AIpom achieved second place in the SemEval-2024 Task 8, demonstrating the effectiveness of this hybrid approach.
What Problem Does This Paper Address?
The rapid advancement of large language models (LLMs) has made it increasingly difficult to distinguish between human-written and AI-generated text. This poses significant challenges, from the potential misuse of LLMs for malicious purposes like generating fake news or academic dishonesty, to the need for reliable methods to identify AI-produced content in various applications. SemEval-2024 Task 8 specifically targets the problem of "Human-Machine Mixed Text Detection," aiming to pinpoint the exact transition point where AI-generated content begins within a larger text. This is a more nuanced task than simply classifying a text as entirely human or AI-generated, and it has direct implications for applications like content moderation, plagiarism detection, and understanding the evolving landscape of digital communication.
Methods
AIpom employs a unique two-stage pipeline that leverages both decoder-only and encoder-only language models.
- Decoder-based Prediction: A decoder-only language model (Mistral-7B-OpenOrca) is fine-tuned to predict the boundary. The fine-tuning prompt instructs the model to output only the machine-generated portion of the text. This approach was found to yield better results than other prompt variations. The decoder's output is then used to insert a special
<BREAK>token into the original text, marking the predicted transition point. - Encoder-based Refinement: Two encoder-only models (DeBERTaV3-Large) are then used to refine the boundary prediction.
- The first encoder is fine-tuned on texts where the
<BREAK>token has been inserted based on the decoder's predictions. - The second encoder is fine-tuned on a mixture of the original training data and the texts with decoder-predicted change points.
Finally, the predictions from the two encoders are aggregated by averaging to produce the final boundary prediction. This pipelined approach allows the encoder models to benefit from the initial predictions of the decoder, leading to improved performance.
Datasets
The primary dataset used in this work is the M4 corpus, specifically for SemEval-2024 Task 8, Subtask C. This corpus contains human-written and machine-generated texts in six languages (English, Chinese, Russian, Urdu, Indonesian, and Arabic) across various domains. The generative models used to create the M4 corpus include OpenAI's ChatGPT and text-davinci-003, LLaMA-1, FLAN-T5, Cohere, Dolly-v2, and BLOOMZ. The dataset is divided into training, development, and official evaluation sets.
Key Insights and Ideas
- Pipelining is Key: The most significant insight is that a pipeline combining decoder and encoder models significantly outperforms using either type of model in isolation. This suggests that the initial "guess" from a generative model can provide valuable context for a discriminative model.
- Decoder's Generative Power for Boundary Detection: Fine-tuning a decoder-only model to generate the AI portion of the text, rather than just classify, proved to be an effective strategy for initial boundary prediction.
- Encoder's Refinement Capability: Encoder models, when fine-tuned on the decoder's predictions and augmented with a special
<BREAK>token, are adept at precisely locating the boundary. - Domain Shift Challenges: The paper highlights a notable performance drop between the development and official test sets, indicating a domain shift issue. This suggests that models trained on one distribution of text might struggle when encountering unseen domains or generators.
Key Results
- AIpom achieved second place out of 33 participating teams in SemEval-2024 Task 8, Subtask C, with a Mean Absolute Error (MAE) of 15.94 on the official evaluation set.
- After the official evaluation, the authors developed an improved solution with an MAE of 15.21.
- Ablation studies confirmed that the combined decoder-encoder pipeline yielded better results than individual models. For instance, fine-tuning only a single encoder model resulted in a higher MAE of 19.97 compared to the pipeline's 15.94.
- The zero-shot performance of the decoder model was poor (MAE of 80.81), but LoRA tuning improved it significantly to an MAE of 17.00.
Limitations and Future Work
- Robustness to Domain Shift: The authors acknowledge that AIpom's robustness needs improvement, particularly when dealing with out-of-domain data. The performance drop between development and test sets underscores this limitation. Future work will focus on enhancing robustness to handle domain-shift scenarios.
- Pipeline Length: The current pipeline consists of a decoder and two encoders. The authors suggest exploring longer pipelines with more complex combinations of models as a future direction.
- Model Combinations: Investigating different combinations of decoder and encoder models could further improve performance.
- Generalizability: While the M4 corpus is diverse, further testing on a wider range of AI generators and domains would be beneficial to assess true generalizability.
Summary updated