Enhancing the healthcare retrieval with a self-adaptive saturated density function

The proximity based information retrieval models usually use the same pre-define density function for all of terms in the collection to estimate their influence distribution. In healthcare domain, however, different terms in the same document have different influence distributions, the same term in different documents also has different influence distributions, and the pre-defined density function may not completely match the terms’ actual influence distributions. In this paper, we define a saturated density function to measure the best suitable density function that fits the given term’s influence distribution, and propose a self-adaptive approach on saturated density function building for each term in various circumstance. Particularly, our approach utilizing Gamma process is an unsupervised model with no requirements for external resources. Then, we construct a density based weighting method for the purpose of evaluating the effectiveness of our approach. Finally, we conduct our experiment on five standard CLEF and TREC datasets, and the experimental results show that our approach is promising and outperforms the pre-defined density functions in healthcare retrieval.