Selecting good expansion terms based on similarity kernel function

In this paper, we propose a novel expansion terms selection model, in which similarity kernel function is used to estimate the relevance between query and candidate expansion terms. In previous method, expansion terms are usually selected by counting term co-occurrences in the documents. However, term co-occurrences are not always a good indicator for relevance, whereas some are background terms of the whole collection. In order to select good expansion terms, similarity kernel function is used in our model to estimate the relevance weight between query and its relevant term extracted from the top-ranked documents in initial retrieval results. The estimated relevance weights are used to select good expansion terms for second retrieval. The experiments on the two test collections show that our expansion terms selection model is more effective than the standard Rocchio expansion.