[Paper Review] Towards Automatic Boundary Detection for Human-AI Collaborative Hybrid Essay in Education

Zijie Zeng, Lele Sha+4 more|arXiv (Cornell University)|Jul 23, 2023

Topic Modeling4 citations

TL;DR

This paper proposes TriBERT, a two-step method for automatic boundary detection in human-AI collaborative hybrid essays, where transitions between human-written and AI-generated content are identified by clustering sentence embeddings and detecting maximal dissimilarity between adjacent prototypes. The approach achieves up to a 22% improvement over the best baseline in in-domain evaluation and 18% in out-of-domain evaluation when using a large prototype size for single-boundary texts.

ABSTRACT

The recent large language models (LLMs), e.g., ChatGPT, have been able to generate human-like and fluent responses when provided with specific instructions. While admitting the convenience brought by technological advancement, educators also have concerns that students might leverage LLMs to complete their writing assignments and pass them off as their original work. Although many AI content detection studies have been conducted as a result of such concerns, most of these prior studies modeled AI content detection as a classification problem, assuming that a text is either entirely human-written or entirely AI-generated. In this study, we investigated AI content detection in a rarely explored yet realistic setting where the text to be detected is collaboratively written by human and generative LLMs (i.e., hybrid text). We first formalized the detection task as identifying the transition points between human-written content and AI-generated content from a given hybrid text (boundary detection). Then we proposed a two-step approach where we (1) separated AI-generated content from human-written content during the encoder training process; and (2) calculated the distances between every two adjacent prototypes and assumed that the boundaries exist between the two adjacent prototypes that have the furthest distance from each other. Through extensive experiments, we observed the following main findings: (1) the proposed approach consistently outperformed the baseline methods across different experiment settings; (2) the encoder training process can significantly boost the performance of the proposed approach; (3) when detecting boundaries for single-boundary hybrid essays, the proposed approach could be enhanced by adopting a relatively large prototype size, leading to a 22% improvement in the In-Domain evaluation and an 18% improvement in the Out-of-Domain evaluation.

Motivation & Objective

• To address the gap in AI content detection for hybrid texts, which are collaboratively written by humans and large language models (LLMs), rather than assuming texts are purely human or AI-generated.
• To formalize the problem of AI content detection as boundary detection—identifying transition points between human-written and AI-generated segments in hybrid essays.
• To develop a method that enables precise localization of AI-generated content, supporting educators in identifying suspicious parts for feedback or revision.
• To investigate the impact of prototype size and boundary count on detection performance, offering practical guidance for real-world deployment.

Proposed method

• The method constructs a hybrid essay dataset by randomly removing sentences from original student essays and replacing them with LLM-generated content using ChatGPT.
• A two-step approach is proposed: first, a TriBERT encoder is fine-tuned to separate AI-generated and human-written content during training by leveraging contrastive learning.
• Sentences in the hybrid text are grouped into prototypes based on consecutive sentence sequences, with each prototype represented as the mean embedding of its constituent sentences.
• The distance between adjacent prototypes in the embedding space is computed, and boundaries are identified as the pairs of prototypes with the maximum dissimilarity.
• The method dynamically adjusts the prototype size based on the number of boundaries: larger sizes are optimal for texts with fewer boundaries (e.g., one boundary), while smaller sizes are better for texts with multiple boundaries.
• The approach is evaluated under both in-domain and out-of-domain settings, comparing performance against fine-tuned BERT and a commercial detector (GPTZero).

Experimental results

Research questions

• RQ1 How can we detect the precise boundaries between human-written and AI-generated content in hybrid essays, rather than classifying an entire text as one or the other?
• RQ2 What is the impact of prototype size on boundary detection performance, especially when the number of boundaries varies?
• RQ3 How does the proposed method compare to existing baselines, including fine-tuned BERT and commercial detectors like GPTZero, in both in-domain and out-of-domain settings?
• RQ4 Can the training process that separates AI and human content during encoder learning significantly improve detection performance?
• RQ5 Under what conditions is the use of larger prototype sizes beneficial for boundary detection in hybrid texts?

Key findings

• The proposed TriBERT method consistently outperformed baseline methods, including fine-tuned BERT and the commercial GPTZero detector, across all experimental settings.
• The encoder training process, which separates AI and human content during fine-tuning, significantly boosted detection performance, indicating the importance of representation learning.
• For single-boundary hybrid essays, using a larger prototype size (p=4) led to a 22% improvement in in-domain evaluation and an 18% improvement in out-of-domain evaluation compared to the best baseline.
• When the number of boundaries increased (e.g., multiple transitions), a smaller prototype size was more effective, suggesting that prototype size should be adapted based on the complexity of the hybrid text.
• The method successfully identified transition points in hybrid texts where commercial detectors like GPTZero failed, even when the overall text was classified with low confidence (e.g., 51% human-written).
• The results demonstrate that TriBERT can serve as a reliable supplementary tool for educators to detect and localize AI-generated content in student essays, enabling targeted feedback or revision.