Evaluating a novel kind of retrieval models based on relevance decision making in a relevance feedback environment

This paper presents the results of our participation in the relevance feedback track using our novel retrieval models. These models simulate human relevance decision-making. For each document location of a query term, information from its document-context at that location determines the relevance decision outcomes there. The relevance values for all documents locations of all query terms in the same document are combined to form the final relevance value for that document. Two probabilistic models are developed, and one of them is directly related to the TF-IDF term weights. Our initial retrieval is a passage-based retrieval. Passage scores of the same document are combined by the Dombi fuzzy disjunction operator. Later, we found that the Markov random field (MRF) model produces better results than our initial retrieval system (without relevance information). If we apply our novel retrieval models using the initial retrieval list of the MRF model, the retrieval effectiveness of our models will be improved. These informal run results using the MRF model used in conjunction with our novel models are also presented.