Insights into Classifying and Mitigating LLMs' Hallucinations

Alessandro Bruno,Pier Luigi Mazzeo,Aladine Chetouani,Marouane Tliba,Mohamed Amine Kerkouri

DOI: https://doi.org/10.48550/arXiv.2311.08117

2023-11-14

Abstract:The widespread adoption of large language models (LLMs) across diverse AI applications is proof of the outstanding achievements obtained in several tasks, such as text mining, text generation, and question answering. However, LLMs are not exempt from drawbacks. One of the most concerning aspects regards the emerging problematic phenomena known as "Hallucinations". They manifest in text generation systems, particularly in question-answering systems reliant on LLMs, potentially resulting in false or misleading information propagation. This paper delves into the underlying causes of AI hallucination and elucidates its significance in artificial intelligence. In particular, Hallucination classification is tackled over several tasks (Machine Translation, Question and Answer, Dialog Systems, Summarisation Systems, Knowledge Graph with LLMs, and Visual Question Answer). Additionally, we explore potential strategies to mitigate hallucinations, aiming to enhance the overall reliability of LLMs. Our research addresses this critical issue within the HeReFaNMi (Health-Related Fake News Mitigation) project, generously supported by NGI Search, dedicated to combating Health-Related Fake News dissemination on the Internet. This endeavour represents a concerted effort to safeguard the integrity of information dissemination in an age of evolving AI technologies.

Computation and Language

What problem does this paper attempt to address?

The paper attempts to address the issue of "hallucinations" generated by large language models (LLMs) when producing text. Specifically, these hallucinations can lead to the generation of incorrect or misleading information, especially in tasks such as question answering systems, machine translation, dialogue systems, summarization, knowledge graph generation, and visual question answering. The main objectives of the paper are: 1. **In-depth analysis of hallucination phenomena**: Investigate the causes of hallucinations and their significance in artificial intelligence. 2. **Classification of hallucination types**: Categorize hallucinations in different tasks, including machine translation, question answering systems, dialogue systems, summarization systems, knowledge graph-based text generation, and visual question answering. 3. **Propose mitigation strategies**: Explore potential strategies to alleviate hallucinations to improve the overall reliability of LLMs. Through this research, the paper aims to provide a theoretical foundation and practical guidance for addressing the hallucination problem, particularly in the HeReFaNMi (Health-Related Fake News Mitigation) project, which is dedicated to reducing the spread of health-related fake news on the internet.

Insights into Classifying and Mitigating LLMs' Hallucinations

Comprehending and Reducing LLM Hallucinations

Hallucination Detection and Hallucination Mitigation: An Investigation

Redefining "Hallucination" in LLMs: Towards a psychology-informed framework for mitigating misinformation

A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models

Developing a Reliable, General-Purpose Hallucination Detection and Mitigation Service: Insights and Lessons Learned

The Troubling Emergence of Hallucination in Large Language Models -- An Extensive Definition, Quantification, and Prescriptive Remediations

MedHalu: Hallucinations in Responses to Healthcare Queries by Large Language Models

Addressing Hallucinations with RAG and NMISS in Italian Healthcare LLM Chatbots

Towards Mitigating Hallucination in Large Language Models via Self-Reflection

A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

Look Within, Why LLMs Hallucinate: A Causal Perspective

Cost-Effective Hallucination Detection for LLMs

Unravelling the Mysteries of Hallucination in Large Language Models: Strategies for Precision in Artificial Intelligence Language Generation

Banishing LLM Hallucinations Requires Rethinking Generalization

Detecting and Mitigating the Ungrounded Hallucinations in Text Generation by LLMs

A Stitch in Time Saves Nine: Detecting and Mitigating Hallucinations of LLMs by Validating Low-Confidence Generation

Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations

Investigating and Addressing Hallucinations of LLMs in Tasks Involving Negation

The Dawn After the Dark: An Empirical Study on Factuality Hallucination in Large Language Models

Mitigating Entity-Level Hallucination in Large Language Models