HALO: An Ontology for Representing Hallucinations in Generative Models

The paper: http://arxiv.org/abs/2312.05209

## Purpose 
The paper addresses the challenge of hallucinations in generative AI, including LLMs like ChatGPT. It presents the Hallucination Ontology (HALO), a formal, extensible ontology written in OWL, offering support for various types of hallucinations and related metadata. This tool is designed to help better understand and categorize hallucinations in AI systems.

## Methods 
- Development of the HALO ontology using the Linked Open Terms (LOT) methodology.
- Identification of primary hallucination categories through literature review.
- Integration of HALO with terms from existing vocabularies for interoperability.
- Use of Google Forms and Python-based programs for data conversion into RDF/XML format.

## Key Findings 
1. HALO supports six distinct types of hallucinations in LLMs, categorized under Factuality and Faithfulness Hallucinations.
2. The ontology allows for systematic documentation of hallucinations, their metadata, and provenance.
3. HALO is validated using competency questions (CQs) and SPARQL queries to demonstrate its utility in empirical studies of hallucinations.
4. HALO is published under an open license and is accessible for public use and contribution.

## Hallucinations in HALO Ontology

### Factuality Hallucinations
1. **Factual Inconsistency**: 
   - Occurs when LLMs produce responses with contradictions or misinformation about real-world facts.
   - Example: Incorrectly identifying "Yuri Gagarin" as the first person to land on the Moon.

2. **Factual Fabrication**: 
   - Refers to LLMs creating responses with unverifiable supporting evidence.
   - Example: Fabricating narratives about dragons existing on Earth without any empirical evidence.

### Faithfulness Hallucination
1. **Instruction Inconsistency**: 
   - Arises when LLM outputs do not align with the user’s prompt instructions.
   - Example: The user requests a translation, but the LLM performs a question-answering task instead.

2. **Context Inconsistency**: 
   - Occurs when the LLM’s response does not align with the provided context.
   - Example: Providing dinner suggestions when asked about breakfast.

3. **Logical Inconsistency**: 
   - Happens when an LLM’s output contains internal logical contradictions.
   - Example: Correctly performing mathematical steps but concluding with an incorrect final answer.

4. **Confidence Mismatch**: 
   - Involves LLMs displaying a high level of confidence in responses that are factually incorrect or nonsensical.
   - Example: Asserting incorrect facts with high confidence.

## Discussion 
This research provides a crucial tool for understanding and managing hallucinations in AI, an area increasingly important as AI systems like ChatGPT become more sophisticated and widely used. HALO bridges the gap between anecdotal observations and systematic study, offering a structured way to analyze hallucinations in LLMs.

## Critiques 
1. The need for continuous evolution of HALO to capture new types of hallucinations as AI technology progresses.
2. The challenge of ensuring comprehensive coverage of hallucination instances, given the decentralized nature of data sources.

## Tags
#AI #Hallucinations #LLMs #ChatGPT #Ontology #HALO #GenAI.

Leave a Comment