Integrating Probabilistic and Semantic Features for AI-Generated Text Localization

Abstract

The proliferation of Large Language Models (LLMs) has led to an increase in AI-Generated Text (AIGT), raising concerns about factual inaccuracies and disinformation. Existing detection methods typically rely on document-level classification, which often fails to capture localized AI modifications or accurately classify short texts. To overcome these limitations, we propose a Probabilistic-Semantic Feature fusion Locator (PSFL), a novel method integrating probabilistic and semantic features for sentence-level AIGT localization. PSFL employs a global attention-driven adaptor network with majority voting, improving short text detection by capturing long-range dependencies. Experimental results on two datasets demonstrate that PSFL outperforms six baseline models, achieving higher mean average precision across various LLMs.

Ähnliche Arbeiten