HALO: An Ontology for Representing and Categorizing Hallucinations in Large Language Models

TL;DR

This work addresses the lack of a formal vocabulary to represent and analyze hallucinations in large language models by introducing HALO, an extensible OWL ontology with two modules (Hallucination and Metadata) that encodes six hallucination types. HALO is evaluated through competency-question driven SPARQL queries on a real-world dataset of 40 prompts collected from diverse sources and tested across three LLMs, demonstrating the ability to model instances, provenance, and metadata. The contributions include a publicly available HALO ontology under CC-BY 4.0, a companion dataset, and a validation framework aligned with the LOT methodology, enabling systematic, cross-model and longitudinal analyses of hallucinations. The work promises to advance reproducible, interoperable research on GenAI reliability and supports ongoing dataset expansion and ontology refinement as models evolve.

Abstract

Recent progress in generative AI, including large language models (LLMs) like ChatGPT, has opened up significant opportunities in fields ranging from natural language processing to knowledge discovery and data mining. However, there is also a growing awareness that the models can be prone to problems such as making information up or `hallucinations', and faulty reasoning on seemingly simple problems. Because of the popularity of models like ChatGPT, both academic scholars and citizen scientists have documented hallucinations of several different types and severity. Despite this body of work, a formal model for describing and representing these hallucinations (with relevant meta-data) at a fine-grained level, is still lacking. In this paper, we address this gap by presenting the Hallucination Ontology or HALO, a formal, extensible ontology written in OWL that currently offers support for six different types of hallucinations known to arise in LLMs, along with support for provenance and experimental metadata. We also collect and publish a dataset containing hallucinations that we inductively gathered across multiple independent Web sources, and show that HALO can be successfully used to model this dataset and answer competency questions.

Figures (5)

• Figure 1: A visualization of HALO, including the Metadata and Hallucination modules that describe it at the highest level and are further articulated in the main text.
• Figure 2: An illustration of the complete workflow for evaluating HALO. Hallucination instances are collected from multiple sources on the internet, typically as response examples from a single LLM. As described in Section \ref{['dataset']}, each prompt is executed across three different LLMs to complete the dataset, and instances are labeled as hallucinations when identified. Only the instances that exhibit hallucinations are then used to create Knowledge Graph (KG) instances, which are subsequently integrated into HALO for evaluation through SPARQL queries.
• Figure 3: An illustrative example showing BARD's responses that hallucinated when prompted in October 2023, but refused to answer in March 2024.
• Figure 4: An illustrative example of a hallucination modeled by HALO (in KG manner) using one data point from the dataset.
• Figure 5: An example of SPARQL used to validate CQ5