SILTD: Structural Information for LLM-generated Text Detection

Abstract The rapid development of large language models (LLMs) has significantly improved the quality and diversity of AI-generated content(AIGC). LLMs-generated text detection plays an important role in preventing the harmful misuse of large language models. Although many detection methods have been proposed, they all have room for improvement and lack generality when detecting multiple models. To tackle these drawbacks, an unsupervised-based Structural Information for LLM-generated Text Detection(SILTD) method is proposed. First, we construct a multi-relational text graph based on the similarity of text features, which aims to model the intricate similarities and correlations between texts in the generative statistical space. Second, we propose a novel unsupervised graph clustering method. The multi-relational graph is transformed into an encoding tree, which is then optimized based on a two-dimensional structure entropy minimization algorithm to achieve hierarchical clustering of texts. Structural entropy minimization enables achieving high-quality clusters, by measuring the uncertainty of random walks within the graph. Finally, we introduce a new method that measures text similarity and computes the intensity of text aggregation within each cluster, to perform in-cluster label inference. Extensive experiments show that, compared to baseline methods, our approach is more effective and generalizable in detecting six popular LLMs across three datasets.