Databases are often considered as the most reliable sources for knowledge extraction. Methods and tools have been proposed to build ontologies from databases, mainly by exploiting relational schemas. However, the resulting ontologies are often far below expectations, especially in terms of expressivity. A significant part of the domain-specific semantics needed to build more expressive ontologies can only be recovered from the stored data. This chapter describes some semantic patterns that are frequently encountered in databases and provides a systematic description of data-driven methods to exploit these patterns for ontology learning.