KSORD (keyword search over relational database) techniques allow users to obtain information from databases, which is just like using search engines. However, the advanced techniques only realize exact queries, but not for fuzzy queries. The Rocchio algorithm of learning classification is introduced which is made a little changed to achieve keyword search over relational databases. According to the dissimilarity and correlation of the quantification calculation between different type objects, returned result sets are ranked in descendant order according to correlation. Thus, the system realizes both exact and fuzzy queries. If users are not satisfied with the initial result sets, they could utilize the Rocchio algorithm to do several relevance feedbacks in order to make results better. We employee the optimal Rocchio algorithm to experiment, the results satisfy the requirements of users. In addition, few non-relevant result sets could improve the performance of searching.