Improving Diversity Of Web Search Results

With the development of the information retrieval technology,people have higher demands for it.Usually,most web retrieval systems only return the search document results according to the relevance of the query and document,and then rank the search results based on the ranking principle PRP(probability ranking principle).The ranked results will be relatively simple in view of the content and cause the redundancy,so could not meet the diversity search needing of users.Moreover,because of the query ambiguity,the single content of the ranked results will cause that the users cannot find any information and give up this search.In information retrieval,the search diversity as one way to solve this problem was paid more and more attention,which aimed at producing the ranked results with the purpose of maximizing the number of the query subtopics covered by the top N documents and minimizing the redundancy.Up to now,many heuristic diversity models have been proposed,but limited features have been considered and extensive works have been needed to tune the model parameters.In our paper,we build the learning model to solve above problems.Specifically,we firstly analyze the heuristic diversity model--Affinity Ranking model to provide the motivation for designing the Learning Affinity Ranking model.And then we design the model learning algorithm by considering the diversity evaluation metrics in the cost function.Through this learning algorithm,we build more effective and robust ranking model than original Affinity Ranking model.In additional,another specific work of our paper is to combine the proposed diversity model to produce more effective diversity ranking model,and then propose the Document Repulsion Theory(DRT)to design the learning algorithm.Two experiments are conducted on diversity task of TREC Web track(2009,2010 and 2011).The results demonstrate that our models significantly outperform a number of state-of-the-art baselines in terms of effectiveness and robustness.