Cross-media topic mining on wikipedia

Abstract

As a collaborative wiki-based encyclopedia, Wikipedia provides a huge amount of articles of various categories. In addition to their text corpus, Wikipedia also contains plenty of images which makes the articles more intuitive for readers to understand. To better organize these visual and textual data, one promising area of research is to jointly model the embedding topics across multi-modal data (i.e, cross-media) from Wikipedia. In this work, we propose to learn the projection matrices that map the data from heterogeneous feature spaces into a unified latent topic space. Different from previous approaches, by imposing the l1 regularizers to the projection matrices, only a small number of relevant visual/textual words are associated with each topic, which makes our model more interpretable and robust. Furthermore, the correlations of Wikipedia data in different modalities are explicitly considered in our model. The effectiveness of the proposed topic extraction algorithm is verified by several experiments conducted on real Wikipedia datasets.