AutoHall: Automated Hallucination Dataset Generation for Large Language Models

Original Paper: https://arxiv.org/abs/2310.00259

By: Zouying Cao, Yifei Yang, Hai Zhao

Abstract:

While Large language models (LLMs) have garnered widespread applications across various domains due to their powerful language understanding and generation capabilities, the detection of non-factual or hallucinatory content generated by LLMs remains scarce. Currently, one significant challenge in hallucination detection is the laborious task of time-consuming and expensive manual annotation of the hallucinatory generation. To address this issue, this paper first introduces a method for automatically constructing model-specific hallucination datasets based on existing fact-checking datasets called AutoHall. Furthermore, we propose a zero-resource and black-box hallucination detection method based on self-contradiction. We conduct experiments towards prevalent open-/closed-source LLMs, achieving superior hallucination detection performance compared to extant baselines. Moreover, our experiments reveal variations in hallucination proportions and types among different models.

Summary Notes

Enhancing AI Reliability: Addressing Hallucinations in Language Models

In the world of artificial intelligence, Large Language Models (LLMs) such as ChatGPT and GPT-4 are leading the way in natural language processing. These models have revolutionized tasks like customer service automation and data analysis.

Despite their advancements, they face a significant challenge: producing hallucinatory content—false or ungrounded information. This issue compromises user trust and the reliability of AI decisions.

Understanding the Hallucination Issue

Hallucinations in LLMs are a widespread issue, stemming from their training on large, unchecked datasets. This leads to coherent but incorrect responses.

The problem is critical in sectors like healthcare and finance, where accuracy is crucial.

Traditionally, detecting hallucinations in LLM outputs has required manual annotation, a costly and model-specific method that lacks scalability and adaptability.

Introducing AutoHall: Advancing Towards Reliable AI

Researchers Cao Zouying, Yang Yifei, and Zhao Hai from Shanghai Jiao Tong University developed AutoHall, an innovative approach to combat hallucinations in LLMs.

AutoHall automates the creation of hallucination datasets for models, offering a scalable, efficient, and model-agnostic detection method based on self-contradiction.

Features of AutoHall

• Automated Dataset Creation: Utilizes fact-checking datasets to produce references for evaluating the hallucinatory content, eliminating manual annotation's need.
• Self-Contradiction Detection: Identifies hallucinations by analyzing contradictions in multiple references for a single claim, avoiding the need for external knowledge bases.

Benefits for AI Engineers

• Scalability: Facilitates the quick generation of large, model-specific datasets for hallucination detection.
• Efficiency: Reduces the resources needed for manual dataset annotation.
• Improved Model Reliability: Helps ensure LLM outputs in enterprise applications are accurate and reliable.

AutoHall in Practice: Boosting AI Trustworthiness

Testing on advanced LLMs like ChatGPT and Llama-2 showed AutoHall could detect hallucinations more accurately than current methods, enhancing LLM reliability across various applications.

Towards a Future Without Hallucinations

AutoHall represents a significant step towards creating more reliable and trustworthy AI by offering an efficient solution to LLM hallucinations.

It promises considerable benefits for AI engineers by providing a tool for developing dependable AI systems, crucial for critical domain applications.

The journey to tackle hallucinations in LLMs continues, but AutoHall equips AI engineers with a powerful tool for ensuring a reliable AI future.

The work by Cao Zouying, Yang Yifei, and Zhao Hai significantly contributes to the ongoing efforts to build AI systems that are both intelligent and reality-based, marking progress in the fight against LLM hallucinations.