FactAlign: Fact-Level Hallucination Detection and Classification Through Knowledge Graph Alignment
Mohamed Rashad, Ahmed Zahran, Abanoub Amin, Amr Abdelaal, Mohamed Altantawy · 2024
Generative Large Language Models (LLMs) have garnered significant attention for their ability to generate human-like text across diverse domains.However, a major obstacle preventing their widespread adoption in production environments is their propensity for 'hallucinations' -the generation of non-factual statements that can erode confidence in their output.Existing hallucination detection approaches either require access to the categorical distribution of the output or rely on external databases to retrieve evidence about generated output.An alternative strategy employs sampling-based techniques, which generate responses multiple times to identify hallucinations.This paper proposes a novel black-box approach to automatically detect and classify hallucinations at a fact level by transforming the problem into a knowledge graph alignment task.This approach, unique in its applications, also allows us to classify detected hallucinations as either intrinsic or extrinsic.Our methodology was evaluated on the WikiBio GPT-3 hallucination dataset for hallucination detection and the XSum hallucination annotations dataset for hallucination classification.Our method achieved a 0.889 F1 for the hallucination detection and 0.825 F1 for the hallucination type classification, without any further training, fine-tuning, or producing multiple samples of the LLM response.