Determining if a claim has been accepted based on the arguments given by judges is an important task for analyzing Law. Applying recent machine learning techniques to automatize this task is however challenging because of the difficulty to obtain labeled dataset in the legal domain - datasets are indeed most often rare, small and expensive in that domain. This article introduces a deep learning model and a methodology to tackle classification tasks in Natural Language Processing when only few labeled examples are available. We show in particular that combining one-shot learning with memory-augmented recurrent neural networks enabled obtaining efficient classification models in such constraining supervised settings. Experimental results and empirical evaluations are proposed using different approaches to represent sentences dealing with several categories of claims expressed in French courts.