A Note on Accurate Extraction of Concept Subsumption Relationships Using Tagged Images

Abstract

This paper presents a method for accurate extraction of concept relationships using tagged images. Previous methods extract concept relationships using either or both of visual features and textual features extracted from tagged images. In the method that we have previously proposed, visual similarity and textual similarity are calculated based on kernel density estimation and word2vec, respectively. Although kernel density estimation considers distributions of the visual features, there is still room for accuracy improvement of concept relationship extraction. In this paper, we utilize locality-constraint linear coding (LLC) to achieve accurate extraction of concept relationships, which is robust to visual variations. The proposed method also utilizes GloVe, which reportedly represents concepts more effectively than word2vec in the field of natural language processing. Experimental results show that LLC and GloVe contribute to effective representation of concepts and improve the accuracy of the subsequent extraction of the concept relationships