Database Selection Using Actual Physical and Acquired Logical Collection Resources in a Massive Domain-specific Operational Environment

摘要

The continued growth of very large data environments such as Westlaw, Dialog, and the World Wide Web, increases the importance of effective and efficient database selection and searching. Recent research has focused on autonomous and automatic collection selection, searching, and results merging in distributed environments. These studies often rely on TREC data and queries for experimentation. We have extended this work to West's online production environment where thousands of legal, financial and news databases are accessed by up to a quarter-million professional users each day. Using the WIN natural language search engine, a cousin to UMass's IN-QUERY, along with a collection retrieval inference network (CORI) to provide database scoring, we examine the effect that a set of optimized parameters has on database selection performance. We also compare current language modeling techniques to this approach. Traditionally, West's information has been structured over 15,000 online databases, representing roughly 6 terabytes of textual data. Given the expense of running global searches in this environment, it is usually not practical to perform full document retrieval over the entire collection. It is therefore necessary to create a new infrastructure to support automatic database selection in the service of broader searching. In this research, we represent our operational environment in two distinct ways. First, we characterize the underlying physical databases that serve as a foundation for the entire Westlaw search system. Second, we create a rearchitected set of logical document collections that corresponds to classes of high level organizational concepts such as jurisdiction, practice area, and document-type. Keeping the end-user in mind, we focus on performance issues relating to optimal database selection, where domain experts have provided complete pre-hoc relevance judgments for collections characterized under each of our physical and logical database models.