Ontology has be applied in many fields, such as data integration, system interoperability, etc. At present, there are many ontology extraction methods based on relational database, but they all assume that the database schema is at least in third normal form (3NF). A set of improved rules to extract ontology is presented in this paper. The ontology generated is described with the OWL language. Compared with other existing methods, this approach can be able to identify a variety of strategies about non-normalized design of relational databases and deal with them, while none of existing methods can identify all of them. The results show that our method provides a more accurate process of ontology learning from non-normalized relational database.