Massive collections of scientific publications are now available on-line thanks to multiple public platforms. These databases usually cover large-scale scientific production over several decades and for a broad range of thematic areas. Today researchers are used to perform queries on these databases with keywords or combination of keywords in order to find articles associated to a precise scientific field. This full text indexation performed for millions of articles represents a huge amount of public information. But instead of being used to characterize articles, can we revert the standpoint and use this information to characterize concepts neighborhood and their evolution? In this paper we give a yes answer to this question looking more precisely at the way concepts can be dynamically clustered to shed light on the way paradigm are structured. We define an asymmetric paradigmatic proximity between concepts which provide hierarchical structure to the scientific database upon which we test our methods. We also propose overlapping categorization to describe paradigms as sets of concepts that may have several usages.