We have restored, refined, and updated this tutorial and added some historical background.
Abstract
This is a light tutorial on OKAPI BM25, a Best Match model where local weights are computed as parameterized frequencies and global weights as RSJ weights. Local weights are based on a 2-Poison model and the verbosity and scope hypotheses and global weights on the Robertson-Spärck-Jones Probabilistic Model.
Introduction
In the early 80s Gillian Venner, Nathalie Mitev, and Stephen Walker (1985, 1987) conducted research work that led to the design and evaluation of online public access catalogs (OPACs) at Polytechnic of Central London (PCL).
The project initial phases spanned from November 1982 to May 1985. The prototype was named OKAPI (Online Keyword Access to Public Information). As Mitev (1985) wrote:
“Designing an online public access catalogue [OPAC]: Okapi, a catalogue on a local area network [LAN] is the final report of a two-year research project ”Microprocessor networking in libraries” which was funded by the British Library and the Department of Trade and Industry, and based at the Polytechnic of Central London.”
“The aim was to produce an OPAC on a LAN, that would be readily usable without training or experience, without sacrificing effectiveness or being tedious for experienced users.”
“The result was a functioning prototype OPAC called Okapi, which has a number of distinctive features: use is eased by coloured keys and a lack of jargon; the system uses search decision trees to select a suitable action at each stage of a search, and it performs automatic Boolean and hyper-Boolean functions where appropriate. The OPAC was installed and evaluated in one of the Polytechnic site libraries.”
Want more? Read the tutorial at
http://www.minerazzi.com/tutorials/okapi-bm25-model.pdf
