[3Xin2] Poster session 1

[3Xin2-56] Investigation of Hallucination Detection Methods in Black Box Large Language Models

〇Asuka Yamazato¹, Kohei Koyama¹ (1.ARISE analytics. inc)

Keywords:Large Lunguage Model, Hallucination, Generative AI, sentence embeddings

Since the launch of ChatGPT, a conversational AI service by OpenAI, its underlying technology, generative AI, has been in the spotlight. One of the problems of generative AI is 'hallucination', a phenomenon where the AI generates content that deviates from actual facts. A method named SelfCheckGPT has been proposed to tackle this problem. This method detects hallucination based on the similarity of outputs, with the reasoning that if a Large Language Model (LLM) knows a given concept well, the responses will likely be similar and contain consistent facts, even when the same prompt is given multiple times.In this study, to verify the performance of SelfCheckGPT for Japanese, we constructed a quiz question answer dataset in Japanese using gpt-3.5-turbo and conducted experiments. The results showed that the performance significantly degraded for Japanese quiz questions. Analysis suggests that this might be due to the similar sentence structures in each output of gpt-3.5-turbo, which in turn depends on the method used for obtaining sentence embeddings.