Automatic recommendation systems (RS), like search engines, have become a must have tool for any website focused on a specific type of article available in a rich catalog, whether those articles are objects, books, films, pieces of music, information (news) or simply pages (hypertext links). The objective of these systems is to select, in their catalog, the elements most likely to interest a particular user, it is to guide the user in his exploration of the data so that it is relevant information. Since all of the work done on content-based RS has used article keywords to recommend them, this approach is usually not the most accurate based on its results. To improve the results of the recommendations, in this article we will propose a technique based on abstracts of scientific articles, since the global information of the article is present in its abstracts. RS has three approaches in general; collaborative filtering (CF), content-based filtering (FBC) and hybrid filtering, if these techniques are used with artificial neural networks like SOM, the system becomes more efficient. In this article, we will propose a content-based approach to improve the accuracy of the list of recommended scientific articles, it is mainly based on the principles of machine learning algorithms (SOM and K-Means). The approach has been tested on a recognized database, and it has shown its performance.