One of the distinctive features of Information Retrieval systems comparing to Database Management systems, is that they offer better compression for posting lists, resulting in better I/O performance and thus faster query evaluation. In this paper, we introduce database representations of the index that reduce the size (and thus the disk I/Os) of the posting lists. This is not achieved by redesigning the DBMS, but by exploiting the non 1NF features that existing Object-Relational DBM systems (ORDBMS) already offer. Specifically, four different database representations are described and detailed experimental results for one million pages are reported. Three of these representations are one order of magnitude more space efficient and faster (in query evaluation) than the plain relational representation.
Download