Large Scale Biomedical Concept Mapping

In this paper, we report on the application of simple retrieval strategies to biomedical concept mapping. We aim at evaluating the performance of a learning-free system tailored to map large collections of concepts, as they can be found in health sciences. Our system is seen as a solution in those cases where machine learning approaches cannot be applied for scalability or data unavailability reasons. For evaluation purposes, the system uses Medical Subject Headings (MeSH) as collection of concepts, and two different collections of MedLine abstracts are used for tuning and evaluation. Unlike most recent text categorization approaches, our approach relies on the combination of two data-independent classifiers. The first classification module uses a tf.idf (term frequency, inverse document frequency) weighting schema, which has been optimally selected for the task. The second classifier is based on regular variations of the concept list. Results emphasize the importance of distinct strategies for minor and major MeSH terms, and show that precision of the hybrid system is significantly improved compared to each single system. For top returned concepts, the system reaches performances comparable to machine learning systems on the OHSUMED collection, while genericity and scalability issues are clearly in favor of the learning-free approach. We draw conclusion on the importance of hybrids strategies for general key words mapping tasks, and discuss the approach in contrast with systems based on machine learning approaches.