A study of the impact of global statistics in distributed information retrieval

Today’s information retrieval systems have to deal with very large data collections and take a distributed approach to achieve scalable retrieval performance. The most widely used approach, called document-partitioning, is to partition the data among multiple search-nodes, which then index their sub-collection independently and are responsible for scoring documents present in their index, against queries. Most of the famous document scoring functions depend on various global (collection-wide) statistics such as document frequency of terms. However, as search-nodes don’t have access to global-statistics and rely on local (sub-collection-wide) statistics for the purpose of scoring, document-partitioning can result in a degraded retrieval performance. In this thesis, we study the impact of the lack of global-statistics on the retrieval performance of a distributed information retrieval (DIR) system. Our experiments show that the performance, as indicated by multiple measures, degrades as the number of search-nodes are increased. We thus conclude that global-statistics are essential to the retrieval performance in a distributed setup. Finally, we present a novel scheme for lazy and adaptive dissemination of global-statistics in a document-partitioned DIR system.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2012

Author dc:creator

• Sehrawat, Nipun

Contributors dc:contributor

• Zhai, ChengXiang

• Distributed Information Retrieval
• Global Statistics
• Retrieval Performance

Statement dc:rights

• Copyright 2012 Nipun Sehrawat

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/31920

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/31920