Evaluation of Prompt Design for Large Language Models in Generating "Easy Japanese" for Multicultural Society

Abstract

As Japan’s labor force becomes increasingly reliant on foreign workers, linguistic barriers have emerged as a major challenge in workplace safety. This study explores the use of large language models (LLMs) to generate “Easy Japanese” from safety manuals in the manufacturing and agriculture sectors. Focusing on prompt design, we compare Zero-shot and One-shot input formats using GPT-4o and a specialized Easy Japanese model. We evaluate output using three methods: manual evaluation, automatic evaluation, and machine-based self-evaluation using ChatGPT. The results show that both input styles produce usable outputs, with One-shot offering greater accuracy and Zero-shot offering speed and ease of use. These findings highlight the potential of LLMs to support low-cost, multilingual workplace communication and promote multicultural inclusion.