Semi-supervised Graph-based Genre Classification for Web Pages
Noushin Rezapour Asheghi, Katja Markert, Serge Sharoff · 2014
Until now, it is still unclear which set of features produces the best result in automatic genre classification on the web. Therefore, in the first set of experiments, we compared a wide range of contentbased features which are extracted from the data appearing within the web pages. The results show that lexical features such as word unigrams and character n-grams have more discriminative power in genre classification compared to features such as part-of-speech n-grams and text statistics. In a second set of experiments, with the aim of learning from the neighbouring web pages, we investigated the performance of a semi-supervised graphbased model, which is a novel technique in genre classification. The results show that our semi-supervised min-cut algorithm improves the overall genre classification accuracy. However, it seems that some genre classes benefit more from this graph-based model than others.