FactCHD: Benchmarking Fact-Conflicting Hallucination Detection

TL;DR

FactCHD introduces a large-scale, multi-domain benchmark for detecting fact-conflicting hallucinations in LLMs, leveraging diverse factuality patterns and golden evidence chains to assess both verdicts and explanations. The authors pair FactCHD with Truth-Triangulator, a triangulation-based framework that combines a tool-enhanced Truth Seeker and a LoRA-tuned Truth Guardian to improve reliability by cross-referencing independent sources. Key contributions include a scalable data-construction pipeline using knowledge graphs (e.g., Wikidata, PrimeKG) and textual knowledge, and evaluation across zero-shot, in-context, and knowledge-enhanced detection settings, showing that specialized detectors and knowledge integration significantly boost factuality detection and explanation quality. The work demonstrates practical impact by delivering interpretable evaluation metrics (FactCls and ExpMatch) and a robust, cross-domain platform for advancing trustworthy AI in real-world, evidence-supported query–response scenarios, with potential extensions to multi-modal and larger-scale deployments.

Abstract

Despite their impressive generative capabilities, LLMs are hindered by fact-conflicting hallucinations in real-world applications. The accurate identification of hallucinations in texts generated by LLMs, especially in complex inferential scenarios, is a relatively unexplored area. To address this gap, we present FactCHD, a dedicated benchmark designed for the detection of fact-conflicting hallucinations from LLMs. FactCHD features a diverse dataset that spans various factuality patterns, including vanilla, multi-hop, comparison, and set operation. A distinctive element of FactCHD is its integration of fact-based evidence chains, significantly enhancing the depth of evaluating the detectors' explanations. Experiments on different LLMs expose the shortcomings of current approaches in detecting factual errors accurately. Furthermore, we introduce Truth-Triangulator that synthesizes reflective considerations by tool-enhanced ChatGPT and LoRA-tuning based on Llama2, aiming to yield more credible detection through the amalgamation of predictive results and evidence. The benchmark dataset is available at https://github.com/zjunlp/FactCHD.

Figures (6)

• Figure 1: Illustration of fact-conflicting hallucination detection example from FactCHD, where the green part represents factual explanation core (body part) in the chain of evidence.
• Figure 2: Overview of the factuality patterns invloved in our FactCHD.
• Figure 3: Domain distribution of FactCHD, where "-p" and "-d" denote domains derived from Wikipedia and Wikidata, respectively.
• Figure 4: Overview of the construction process of FactCHD.
• Figure 5: Overview Truth-Triangulator. Here we designate the as the "Truth Guardian" based on Llama2-7B-chat-LoRA while as the "Truth Seeker" based on GPT-3.5-turbo (tool) in our experiments. We want the "Fact Verdict Manager" to collect evidence from different viewpoints to enhance the reliability and accuracy of the obtained conclusion.