Recently, there has been considerable progress in the classification of textual documents. However, the models used must generally be trained beforehand with many labelled samples. It is possible to reduce this number of samples in order to perform this task by better selecting the examples to be annotated using active learning techniques. This can reduce the cost of the process by reducing human intervention. In this study, we will adapt recent deep active learning techniques used for image classification to the case of text analysis. In particular, we will be attentive to the contribution of deep active learning depending on the architecture used (LSTM or CNN). We will validate our hypotheses on data sets from the literature.