This paper presents the application of a new method of feature selection for the analysis of the evolution and interaction of scientific domains. The query of bibliographic databases provides a corpus of scientific publications in different fields. Every scientific field is considered as a class issued from a machine learning process, either supervised or unsupervised, and each document is represented by a bag of words. Therefore, it is possible to select the most significant words of every class (domain). We then represent the words-classes relationships by a graph whose edges are weighted by a function of contrast. Thus we highlight the specific words for each area and those that are multidisciplinary. In addition, the joint analysis of two periods of time allows us to appreciate the evolution of the scientific fields.