[1G5-OS-26b] OS-26

[1G5-OS-26b-01] Automatic Construction Method for Daily Life Dataset Using Commonsense Knowledge in Large Language Model

〇Jin Aoyama^1,2, Takeshi Morita^1,2, Takanori Ugai^2,3, Shusaku Egami², Kenichiro Fukuda² (1. Aoyama Gakuin University, 2. National Institute of Advanced Industrial Science and Technology, 3. Fujitsu Limited)

Keywords:Large Language Model, VirtualHome, planning, Knowledge Graph

The Knowledge Graph Reasoning Challenge for Social Issues (KGRC4SI) 2022 was held in Japan. The challenge aimed to encourage the development of systems capable of identifying and explaining dangerous situations that might occur in the homes of older people. To achieve this goal, the organizers provided videos simulating daily activities with the household simulator VirtualHome and knowledge graphs converted from the videos using VirtualHome2KG. The primary task of the KGRC4SI was to identify dangerous situations from the provided video and knowledge graph. However, much data representing daily life situations is needed since not all videos and knowledge graphs include dangerous situations. Creating action scripts manually is not intuitive, and it will be easier if users can communicate their intentions and requirements using natural language statements. We have proposed a system for generating action scripts from descriptions of daily life activities using GPT-3.5 Turbo, one of the LLMs, and evaluated their similarity to correct data and execution rate. This research evaluates the system using GPT-3.5 Turbo, GPT-4, and Llama 2 and compares and discusses the results.