Practical Fact Checking System for LLMs

The use of Large Language Models (LLM) like ChatGPT in real-world product solutions is significantly limited by the well-known issue of hallucinations. Various methods exist in order to overcome this issue automatically, such as using a different LLM to provide feedback on the accuracy of the generated text or by examining the output consistency given multiple sampled responses. However, these approaches do not guarantee factual accuracy, which is crucial in many specialized domains. In order to enhance the factual correctness of generated text by LLM models, a combination of manual annotations and supportive tools are required. We suggest a practical fact checking system tailored specifically for LLMs which combines a hybrid approach (human and machine) to evaluate the correctness of the generated text. This is particularly vital in fields where hallucinations present significant challenges. We use proprietary LLMs, both directly and through Retrieval-Augmented Generation (RAG), to offer users informed feedback on potential hallucinations via a user-friendly interface. We apply our methodology to the task of generating aspect values for video games in listings from an E-commerce marketplace, demonstrating the utility of our approach.