Instance Classification using Co-Occurrences on the Web

We present a novel unsupervised approach to mapping artrelated instances (such as music artists and painters) to subjective categories like genre and style. We base our approach on co-occurrences of the two on the web, found with Google. The co-occurrences are found using three methods: by identifying the search engine counts, by analyzing Google excerpts found by querying patterns and by scanning full documents. Per instance, we use the same co-occurrence-based approach to find its nearest neighbors, i.e. the most related instances. These results can be combined in order to create a more reliable classification. We tested and compared the three methods on two different domains: mapping music artists to genres, and painters to art-styles. The results show that the use of related instances indeed improves the precision of the classification. Moreover, the methods with the lowest Google Complexity perform best.