KG-UQ: Knowledge Graph-Based Uncertainty Quantification for Long Text in Large Language Models

Abstract

With the commercialization of large language models (LLMs) and their integration into daily life, addressing their susceptibility to hallucinations-unfactual information in generated outputs-has become an urgent priority. Existing uncertainty quantification (UQ) methods often rely on access to LLMs' internal states, which is unavailable for closed-source models like GPTs, or are primarily designed for short text. Current research on long text typically evaluates sentences individually, overlooking smaller semantic units that better capture the text's complexity. Recognizing the potential of knowledge graphs (KGs) to extract structured relationships from unstructured text, we propose KG-UQ, a UQ method leveraging KGs to address the semantic intricacies of long text. Our approach involves constructing KGs from long-text outputs and utilizing their embeddings to estimate uncertainties. Through our analysis, we demonstrate that knowledge graphs are an effective tool for decomposing long text into fundamental statements. However, we also highlight the increased uncertainty introduced during KG construction, stemming from inherent challenges in accurately capturing all semantic information.