A Study on Understanding the Meaning of Adverbs Expressing Actions with Spectral Mixture Kernel

Abstract

In recent years, the necessity of using robots at home has been increasing, and it is expected for them to behave sharing the same sense with people in our everyday lives. In order to do so, they need to understand the meaning of vague words typically expressed with adverbs and behave based on understanding. Therefore, we can say that a robot can understand the meaning of adverbs if it can find common features for multiple behaviors for a particular adverb. In this study, we try to understand the meaning of adverbs expressing human motions. We reduce the dimension of human pose data in movies with Gaussian Process Latent Variable Model, and then identify several kernels which compose of the data in each dimension by means of Spectral Mixture Kernel. Through the experiments, we have found common features for the kernels to express human motions of a particular adverb.