Neural networks have become a preferred tool for text classification tasks, demonstrating state of the art performances when trained on a large set of labeled data. However, in an early active learning setup, the scarcity of the ground-truth labels available severely penalizes the generalization capability of the neural network. In order to overcome such limitations, in this paper, we introduce a new learning strategy, which consist of inserting in the early stages of the learning process some additional, local and salient knowledge, presented under the form of simulated, human like rationales. We show how such knowledge can be automatically extracted from documents by analyzing the class activation maps of a convolutional neural network. The experimental results obtained demonstrate that the exploitation of such rationales permits to significantly speed-up the learning process, with a spectacular increase of the accuracy rates, starting from a very reduced number of documents (10–20).