Improving Relevance Feedback in Language Modeling Approach: Maximum a Posteriori Probability Criterion and Three-Component Mixture Model

Abstract

Recently, researchers have tried to extend a language modeling approach to apply relevance feedback. Their approaches can be classified into two categories. One typical approach is the expansion-based feedback that sequentially performs ‘term selection’ and ‘term re-weighting’ separately. Another approach is the model-based feedback that focuses on estimating ‘query language model’, which predicts well users’ information need. This paper improves these two approaches of relevance feedback by using a maximum a posteriori probability criterion, and a three-component mixture model. A maximum a posteriori probability criterion is a criterion for selection of good expansion terms from feedback documents. A three-component mixture model is the method that eliminates the noise of the query language model by adding a ‘document specific topic model’. The experimental results show that our methods increase the precision of relevance feedback for a short length query. In addition, we make some comparative study between several relevance feedbacks in three document collections.