文档库

最新最全的文档下载
当前位置:文档库 > Neural Relevance Feedback for Information Retrieval

Neural Relevance Feedback for Information Retrieval

Neural Relevance Feedback for Information Retrieval

Fabio Crestani

International Computer Science Institute

1947Center Street,Suite600

Berkeley,CA94704,USA

fabioc@http://www.wendangku.net/doc/10bd3970f242336c1eb95e9d.html

Abstract.Relevance feedback is a technique used in interactive Information Re-

trieval(IR)systems to enable a user to provide additional information to help

the system identify more relevant documents.The additional information is pro-

vided in the form of relevance judgements on retrieved documents.One of the

most advanced relevance feedback technique in operative IR system is based on

a probabilistic function.Recent results show that it is possible to implement rel-

evance feedback also using neural networks.This paper presents the results of

an experimental investigation into the use of the Back Propagation learning al-

gorithm for implementing relevance feedback.The investigation compares the

performance of the proposed neural relevance feedback technique against to and

in combination with probabilistic relevance feedback.The results obtained seem

to indicate that,while probabilistic relevance feedback often outperforms neural

relevance feedback,the combination of the two techniques is more effective than

both techniques taken separately.

1Introduction

In the last50years much effort has been devoted to improve the performance of In-formation Retrieval(IR)systems.In recent years research has explored many different directions trying to use results achieved in other areas,like artificial intelligence,neural networks,and expert systems[9].

Previous research shows that,though giving encouraging results,neural networks (NN)cannot be effectively used in IR at the current state of the technology(see[21,8] for extensive reviews of the work done in this area).The scale of real IR applications, where hundreds of thousands of documents are used,makes it impossible to use NN in an effective way,unless we use poor document and query representations.However, recent results[15,5]show that it may be still possible to use NN in IR for very specific tasks where the number of patterns involved(and therefore the training)is reduced to a manageable size.

In this paper we therefore investigate the possibility of using NN in IR and in par-ticular we focus on their use in the process called“relevance feedback”.Relevance feedback is not always present in operative IR systems,but it has been widely recog-nised to improve retrieval performance[14,13,10].The purpose of this research is to Most of the work reported in this paper was done while the author was at the Department of Computing Science of the University of Glasgow,Scotland.