We propose a model that employs convolutional neural networks (CNN) to evaluate sociomoral reasoning maturity, a key social ability, necessary for adaptive social functioning. Our model is used in a serious game to evaluate learners. It uses pre-annotated textual data (verbatims) and a coding scheme (SoMoral) applied by experts in psychology. State of the art text classification algorithms (Support Vector Machine, Naïve Bayes, etc.) achieved low results in our context in contrary to the CNN that achieved best results with little fine tuning on the input data representation. We use a simple but efficient input data vectors representation learnt directly from the dataset without loosing the sentences 'semantic'. We present a series of experiments with 5 baseline text classification algorithms and 4 baseline data representation. The results show that our model can predict the level of sociomoral reasoning with about 92% of accuracy. Our findings allow not only to advance the textmining field but also the user modeling in highly social adaptive systems. [For the full proceedings, see ED596512.]