A Mixture Clustering Model for Pseudo feedback in Information Retrieval

Information Retrieval (IR) refers to retrieving relevant documents from a large document database according to a user-submitted query, and is among the most useful technologies to overcome information overload. For example, Web search engines are now essential tools for everyone to find information on the Web. Indeed, search capabilities are becoming more and more popular in virtually all kinds of information management applications. Given a query, a retrieval system would typically estimate a relevance value for each document w.r.t. this query, and rank the documents in the descending order of relevance. Over the decades, many different retrieval models have been proposed and tested, including vector space models , probabilistic models , and logic-based models [6]. As a special family of probabilistic models, the language modeling approaches have attracted much attention recently due to their statistical foundation and empirical effectiveness [5, 2]. A particular effective retrieval model based on statistical language models is the Kullback-Leibler (KL) divergence unigram retrieval model proposed and studied in [4, 8]. The basic idea of this model is to measure the relevance value of a document w.r.t. a query by the Kullback-Leibler divergence between the corresponding query model and the document model. Thus the retrieval task essentially boils down to estimating a query unigram model and a set of document unigram language models. The retrieval accuracy is largely affected by how good the estimated query and document models are. In this paper, we study how to improve the query model estimation through fitting a mixture model to some number of top ranked documents, which are retrieved by the original query itself. We present a new mixture model that extends and improves an existing mixture feedback model and addresses its two deficiencies. We study parameter estimation for this mixture model, and evaluate the model on a document set with 160, 000 news article documents and 50 queries. The results show that using the new mixture model not only