This dissertation introduces the problem of content- based query replacement. It motivates the term correlation modeled using MRFs and the sampling based technique to learn the search phrase definitions. These definitions are used as alternative queries to achieve higher accuracy in several retrieval tasks. Retrieving relevant documents while eliminating irrelevant documents for an user's query is a challenging task which involves a good understanding of the relation between the data and the query as well as developing algorithms that can efficiently measure the relevance of the data to the query. As part of this dissertation, we have developed a hypothesis to reduce the problem of mining search phrase definitions significantly by modeling the joint distribution of terms as a product of conditional distributions, modeled as a Markov Random Field. We assume that there exists an underlying joint distribution among terms that are used to describe the search phrase. The modeling we propose is a condensed representation of inter- term relationship and it appears to capture insight statistics among terms. describes a target phrase.