Online Classification with Partially Labelled Texts

In this investigation, we propose a novel approach to document stream classification using both online topic model and partially labelled documents. Although we may have several features for the classification, it seems natural that these features may vary dynamically depending upon the contents of stream. This is because they depend heavily on each theme within one class while we should follow dynamic mixture of them. Especially in the stream of news articles, word frequency changes dramatically because of bursts of the themes. Here we propose a dynamically learning method based on topic models assuming prior distribution of probabilities over classes adjusted by partially labelled documents in stream.