Clustering methods are nowadays commonly used to improve knowledge base. The main objective is to gather objects based on similarity between their characteristics. However, resulting clusters usually lack of semantic due to too macroscopic insights. This is due to the fact that features are mostly considered as independent dimensions of the metric space in the clustering process. The dependency relationship can be represented by a partial order between features to manage the abstraction and a relation of type is-a can bind them. Therefore, using domain taxonomy in the clustering process, i.e. prior knowledge, should bring more semantic information. More specifically, we suggest a potential alternative based on semantic similarity measures to the use of classical distances within the clustering process. Such semantic measures have the property to take into account relationships between features to address correlation issue. Experiments have been conducted on a real dataset from the biomedical field. Results underline the interest of our proposal.