[2M4-OS-3a] OS-3 (1)

[2M4-OS-3a-02] Learning of Relative Spatial Concepts from Spoken User Utterances Using Mixture Distribution

〇Rikunari Sagara¹, Ryo Taguchi¹ (1. Nagoya Institute of Technology)

Keywords:Relative concepts, Unsupervised learning, Language acquisition

This paper presents an improved method for learning relative spatial concepts using mixture distribution. Service robots are required to learn and understand relative spatial concepts used in our daily life. Our proposed method enables a robot to learn the concepts and phoneme sequences which represent the concepts from utterances without any prior knowledge of words. In the generative model of the proposed method, mixture von Mises distribution is used for generating a relative angle. This enables the robot to learn relative spatial concepts which are separated into several parts. The experimental result showed that the concept “yoko”, which means “side” in English, learned correctly by proposed method. Moreover, syllable sequences representing the concepts were learned correctly.