This paper describes a method aiming at the extraction of generalization/specialization hierarchies contained in a relational database. This reverse engineering approach takes advantage of two major characteristics: first, we use DDL and DML specifications as well as data in a combined way, secondly, we provide not only generalization/specialization hierarchies but also integrity constraints allowing us to elicit the generalization/specialization links hidden in the structures and instances of the database. The result of the process consists of an enriched conceptual representation of the relational database. This approach is mainly based on heuristics. The heuristic rules map a relational meta-model onto a conceptual one. They are divided into three categories: semantics suspicion rules, reinforcement rules and confirmation rules. We illustrate our approach using a fairly complex example. Some extensions are discussed.