Learning of Abstract Concepts and Words Based on Model Structure Selection Using Variational Bayes

Shinya NAKAMURA  Takayuki NAGAI  Naoto IWAHASHI  Ken SATOH  Hideki ASOH 

Publication
D - Abstracts of IEICE TRANSACTIONS on Information and Systems (Japanese Edition)  Vol.J92-D  No.4  pp.467-479
Publication Date: 2009/04/01
Online ISSN: 1881-0225
Print ISSN: 1880-4535
Type of Manuscript: PAPER
Category: 
Keyword: 
abstract concept,  model selection,  variational Bayes,  language acquisition, 

Full Text(in Japanese): PDF(574.4KB)

Summary: 
In this paper, a learning algorithm of abstract concepts and words is proposed. The higher level abstract concepts are formed by combining visual information of an object and possible action through Human-Robot interaction. Such higher level abstract concepts play a very important role in our every day communications. This paper deals with abstract concepts and words indicating functional categories, which are based on possible action to some specific objects such as "edible". The proposed algorithm is based on the graphical model, which previously encodes the relationships between objects and possible actions. In order to learn a new word, multiple candidates' graphical models, each of which represents a possible interpretation of the word, are generated from a basic graphical model. Then the learning of the new word becomes the problem of optimal model selection based on the Bayesian criterion. The variational Bayes method is involved in the estimation of model parameters, since it makes the structure learning possible. The validity of the proposed algorithm is shown through some experiments under various conditions.