External methods to address limitations of using global information on the narrow-down approach for hierarchical text classification

Classifying documents to a large-scale web taxonomy is a challenging research problem because of a large number of categories and associated documents in the taxonomy. The state-of-the-art solution known as the narrow-down approach utilizes a search engine to reduce an entire category hierarchy to most relevant categories and selects the best one among them using a classifier. In a recent language modelling approach, top-level category information (or global information) was used in judging the appropriateness of a local category, which led to performance improvements. However, we observe that using global information has a limited influence on the final category selection under some conditions. First, global information may be inaccurate even though it is generated by a top-level category classifier using an entire hierarchy. Second, it has little influence when two competing categories share the same top-level category or when the local category information has too strong an influence on the final category selection. To resolve the limitations, in this paper, we propose two external methods: (1) a meta-classifier with novel dependency features among top-level categories based on an ensemble learning framework; and (2) a query modification model based on a statistical feedback method to improve the query document representation instead of just juggling with information in the hierarchy. Our methods were evaluated using the Open Directory Project test collection.