An Online Analytical System for Multi-Tagged Document Collections

Abstract

The New York Times Annotated Corpus and the ACM Digital Library are two prototypical examples of document collections in which each document is tagged with keywords and significant phrases. Such collections can be viewed as high-dimensional document cubes against which browsers and search systems can be applied in a manner similar to online analytical processing against data cubes. The tagging patterns in these collections are examined and a generative tagging model is developed that can mimic the tag assignments observed in those collections. When a user browses the collection by means of a Boolean query over tags, the result is a subset of documents that can be summarized by a centroid derived from their document term vectors. A partial materialization strategy is developed to provide efficient storage and access to centroids for such document subsets. A customized local term vocabulary storage approach is incorporated into the partial materialization to ensure that rich and relevant term vocabulary is available for representing centroids while maintaining a low storage footprint. By adopting this strategy, summary measures dependent on centroids (including bursty terms, or larger sets of indicative documents) can be efficiently and accurately computed for important subsets of documents. The proposed design is evaluated on the two collections along with PubMed (a held-back document collection) and several synthetic collections to validate that it outperforms alternative storage strategies.

Finally, an enhanced faceted browsing system is developed to support users' exploration of large multi-tagged document collections. It provides summary measures of document result sets at each step of navigation through a set of indicative terms and diverse set of documents, as well as information scent that helps to guide users' exploration. These summaries are derived from pre-materialized views that allow for quick calculation of centroids for various result sets. The utility and efficiency of the system is demonstrated on the New York Times Annotated Corpus.