Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling

TL;DR

This paper addresses hallucinations in long-text generation by black-box LLMs under zero-resource constraints. It introduces a graph-based context-aware hallucination detection framework (GCA) that first extracts knowledge triples via triple-oriented segmentation, then models and reason over contextual dependencies with an RGCN, and finally applies reverse verification through three reconstruction tasks. The method yields a composite score that fuses graph-based consistency and triple reconstructions, improving detection accuracy over diverse baselines across multiple datasets. The results demonstrate the practical value of considering inter-triple dependencies and reconstruction-based validation for robust, zero-resource hallucination detection in open-ended text.)

Abstract

LLMs obtain remarkable performance but suffer from hallucinations. Most research on detecting hallucination focuses on the questions with short and concrete correct answers that are easy to check the faithfulness. Hallucination detections for text generation with open-ended answers are more challenging. Some researchers use external knowledge to detect hallucinations in generated texts, but external resources for specific scenarios are hard to access. Recent studies on detecting hallucinations in long text without external resources conduct consistency comparison among multiple sampled outputs. To handle long texts, researchers split long texts into multiple facts and individually compare the consistency of each pairs of facts. However, these methods (1) hardly achieve alignment among multiple facts; (2) overlook dependencies between multiple contextual facts. In this paper, we propose a graph-based context-aware (GCA) hallucination detection for text generations, which aligns knowledge facts and considers the dependencies between contextual knowledge triples in consistency comparison. Particularly, to align multiple facts, we conduct a triple-oriented response segmentation to extract multiple knowledge triples. To model dependencies among contextual knowledge triple (facts), we construct contextual triple into a graph and enhance triples' interactions via message passing and aggregating via RGCN. To avoid the omission of knowledge triples in long text, we conduct a LLM-based reverse verification via reconstructing the knowledge triples. Experiments show that our model enhances hallucination detection and excels all baselines.

Figures (2)

• Figure 1: GCA framework. We extract triples from the original response and sampled responses (left-upper corner). Then, we construct a graph for each response with the extracted triples and perform message passing and aggregation on the graph (as the upper branch). We conduct reverse validation for each part of the triples with three reconstruction tasks (as the lower branch).
• Figure 2: Node distribution comparisons with (red points) and without (blue points) RGCN on the PHD (top) and WikiBio (bottom).