Abstract : We propose a method to improve the performance of Random Forests for classifying short texts interactively. In short text classification, the principle of learning algorithms is to build a static model using a training dataset, then to use this model to classify new texts. Many works concentrate on improving the representation of the data as a way to build better models. We intend to tackle the problem in two ways: first by abstracting data to solve the problem of sparseness, and second by taking benefit from already classified data to continuously improve the model. Besides, in order to alleviate the amount of manual annotation, we propose an interactive method in which a manual correct annotation is required only for some misclassified texts, which are then incorporated into the training data to build an updated model. An important challenge is then to determine when to trigger this operation and how to perform the update. Applied on the standard search-snippets dataset, our method allowed a significant improvement.