Recent years, the availability of public accessible structured resources like XML on the web has led to active developments of structural retrieval systems. With these systems, users will be able to query for information from structured resources on the web efficiently. When querying, it is obvious that usage of structural information in query increases the precision of retrieval system. However, general web users are more familiar with unstructured query such as natural language or keywords, which contains no structural information. This motivates us to find a retrieval method that supports querying which is simpler and familiar to user, i.e. unstructured query, but at the same time, does not overlook the usage of structural information in query. Hence, we propose a solution that automatically adds structural information to the unstructured query, and represents it as a Mediated Query. The mediated query is an intermediate query in structured form to bridge the gap of structural differences between unstructured query and structured resources. As the selection of correct structural information that reflects the query context is crucial for better retrieval performance, we develop a method to obtain this information by learning the semantics of a set of terms extracted from structured resources. The semantics of a term is defined by its concept and context. We represent the term and its semantics using the Semantic Prediction Model. The model will be used in reasoning the context of query and the process of creating mediated query. The mediated query is then matched against structured resources to obtain relevant results.