Dynamic taxonomy composition via keyqueries

This paper presents an unsupervised framework for dynamic, subject-oriented taxonomy composition in digital libraries, which can naturally integrate existing library classification systems. The taxonomy classes in our approach correspond to so-called keyqueries that are run against the digital library´s full-text retrieval system. Given a document, a keyquery is a set of few keywords for which the document achieves a high relevance score. Keyqueries can hence be viewed as a general and concise description of the returned retrieval results. The keyquery framework addresses important problems of static classification systems: overlarge classes and overly complex taxonomy structures. If, for instance, a leaf class grows to an indigestible size, keyqueries for the contained documents provide a suitable split mechanism. Since queries are well-known to library users from their daily web search experience, they increase the structural complexity in a transparent way. The paper presents also a strategy for taxonomy-based library exploration. Given a user´s information need in the form of library documents, we synthesize a hierarchy of keyqueries that covers this library subset. We manage to solve this difficult set covering problem on-the-fly by combining inverted and reverted indexes along with heuristic search space pruning within a map-reduce application. An empirical evaluation with an ACM collection of scientific papers demonstrates the efficiency and effectiveness of our taxonomy composition framework.