In keyword search over relational databases (KSORD), retrieval of user’s initial query is often unsatisfying. User has to reformulate his query and execute the new query, which costs much time and effort. In this paper, a method of automatically reformulating user queries by relevance feedback is introduced. The method adopts a ranking method based on vector space model to rank KSORD results. Based on the results of user feedback or pseudo feedback, it computes expansion terms based on probability and reformulates the new query using query expansion. Experimental results verify that after KSORD systems executing the new query, more relevant results are presented to user.