[3G1-OS-24a] 日常生活知識とAI

[3G1-OS-24a-03] Detection and Explanation of Dangerous Behaviors in Home using Knowledge Graph and GPT

Reki Asano¹, Mitsuki Hamamichi¹, Kou Mu¹, 〇Takahito Mori¹, Shohei Yano¹, Takeshi Morita¹ (1. Aoyama Gakuin University)

Keywords:Knowledge Graph, Knowledge Graph Embedding, GPT, SPARQL

In order to use AI technology safely and securely in society, AI technology that can be explained is necessary. In the "Knowledge Graph (KG) Reasoning Challenge [Real World Version]" contest about explainable AI, the safety of the elderly was used as a subject, and one of the tasks was to detect, explain, and present solutions to dangerous situations in the home. In addition, in this challenge, the simulation information using the virtual space in the home is converted into KG, and a standardized data set is provided along with the video. In this study, using this dataset, we propose a framework to detect risky behaviors in the home, explain the risks, and present solutions. In the detection of dangerous behavior, inference rules for potentially dangerous behavior such as ``walking near an object'' are defined using SPARQL queries, and behaviors that satisfy the conditions are detected. In addition, we try to detect further dangerous behaviors that could not be obtained by reasoning by embedding KG in the detected behaviors. After that, a template is used to create a natural sentence that asks what the risk and solution of the detected dangerous behavior is, and it is input to GPT-3 and ChatGPT. We evaluate the output results of each GPT and verify the validity of the proposed method.