DEVELOPING A PERSONALIZED ARTICLE RETRIEVAL SYSTEM FOR PUBMED

Abstract

PubMed keyword based search often results in many citations not directly relevant to the

user information need. Personalized Information Retrieval (PIR) systems aim to improve the

quality of the retrieval results by letting the users supply information other than keywords.

Two main problems have been identified for the current PIR systems developed for PubMed:

(1) requiring the user to supply a large number of citations directly relevant to a search topic,

and (2) producing too many search results, with a high percentage being false positives. This

study developed a Personalized Article Retrieval System (PARS) for PubMed to address

these problems. PARS uses two main approaches to find the relevant citations to the given

information need: (1) Extending the PubMed Related Article (PMRA) feature and (2) Text

classification based Multi Stage Filtering (MSF) method. Both approaches require only a

small set of citations from the user, and reduce the search output size by eliminating the

false-positive citations in the search output. PARS has been experimentally evaluated using

the TREC 2005 dataset, and empirically evaluated by subject experts from the biomedicine

field. Results show the PARS system is able to produce retrieval results of better quality

than the existing PIR systems for PubMed.