Overlapping Clustering Method Using Local and Global Importance of Feature Terms at NTCIR-4 WEB Task.

In NTCIR-4 Web Task D (Topical Classification Task), we present an overlapping clustering method for a Japanese meta search engine as an alternative to a list of ranked retrieval results which most search engines adopt to present the retrieval results. The proposed method clusters the retrieval results dynamically according to the following two steps: (1) cluster labels consisting of the most important feature terms extracted from the retrieval results are generated first; then (2) each document is classified into one or more (i.e., overlapping) generated clusters based on its relevance to the feature term. The evaluation results showed that the proposed method in formal run achieved better retrieval effectiveness compared to the average of all the participants in Task D.