Psychomime Classification Using Similarity Measures and Fuzzy c-Means

Abstract

Our aim is to automatically classify psychomimes such as ukiuki and wakuwaku, which appear frequently in Japanese daily life. This task is important because psychomimes represent users' emotions and have multiple meanings with various contexts. Our previous study focused on these characteristics using fuzzy c-means algorithms. However, only one data set tended to be arranged near a centroid, with the other data located away from the centroid of its group. This means it is difficult to regard the second psychomime as belonging to the same group. This arrangement might have resulted from the vector space description, which used 879 content words to form a high-dimensional vector. Therefore, we attempt two solutions to reduce the dimensionality, namely the adoption of similarity descriptions, such as the cosine similarity, and the application of vector quantization algorithms to our data. We conduct two experiments on psychomime classification using these procedures, and compare the results. © Springer-Verlag Berlin Heidelberg 2014.