The anatomy of a Clustering Engine for Web Snippets

Recently there has been a surge of commercial interest about novel IR-tools, like Vivisimo or Groxis, that support the user of a search engine in his/her query formulation and query refinement. The basic idea is that the snippets returned by the search engine are grouped into clusters which are then organized in a hierarchy whose nodes are properly labeled via meaningful sentences. Each sentence must capture the "theme" of the snippets contained into the cluster it labels. This way the user is provided with a small, but intelligible, picture of the query answers at various levels of details. Despite this commercial interest, we found just four scientific papers on this topic. None of them achieved results comparable to Vivisimo, that actually represents the state-of-the-art. In the present paper we address this problem in its full generality: labels of variable length for denoting the clusters, labels drawn from the Web snippets as non contiguous sequences of terms, clusters possibly overlapping and organized within a hierarchy. We achieve this results by means of an algorithmic approach that exploits some innovative ideas, at least from the academic side