With the advent of social networks and the multiplication of product messages about companies, better understanding of customer feedback has become a key issue. Clustering techniques and thematic modeling already allow to observe the main trends observed in this data. It is interesting, from an anticipatory perspective, to observe the emerging themes and to identify them before they grow in size. To solve this problem, we studied the use of LDA models to detect documents related to these emerging themes. We tested three systems on several novelty arrival scenarios in the data stream. We show that the thematic models allow to detect this novelty but that it depends on the scenario considered.