Improving Text Retrieval in Medical Collections Through Automatic Categorization

Abstract

A current and important research issue is the retrieval of relevant medical information. In fact, while the medical knowledge expands at a rate never observed before, its diffusion is slow. One of the main reasons is the difficulty in locating the relevant information in the modern and large medical text collections of today. In this work, we introduce a framework, based on Bayesian networks, that allows combining information derived from the text of the medical documents with information on the diseases related to these documents (obtained from an automatic categorization method). This leads to a new ranking formula which we evaluate using a medical reference collection, the OHSUMED collection. Our results indicate that this combination of evidences might yield considerable gains in retrieval performance. When the queries are strongly related to diseases, these gains might be as high as 84%. This shows that information generated by an automatic categorization procedure can be used effectively to improve the quality of the answers provided by an information retrieval (IR) system specialized in the medical domain.