Abstract
Most relevance discovery models only consider document-level evidence, which may introduce uncertainties to relevance features. Research in information retrieval shows that adopting passage-level (i.e. paragraph-level) evidence can improve the performance of different models in various retrieval tasks. This paper proposes an innovative and effective relevance method based on paragraph evidence to reduce uncertainties in the relevance features discovered by existing models. The method exploits latent topics in the relevance feedback collection to estimate the implicit paragraph relevance. It uses random sets to effectively model the intricate relationships between paragraphs, topics and features to deal with the associated uncertainties. Experiments are conducted using the standard RCV1 dataset, its TREC filtering collections and six popular performance measures. The results confirm that the proposed Uncertainty Reduction (UR) method can significantly enhance the performance of 12 models for relevance feature selection.