Determining if a claim is accepted given judge arguments is an important non-trivial task in court decisions analyses. Application of recent efficient machine learning techniques may however be inappropriate for tackling this problem since, in the Legal domain, labelled datasets are most often small, scarce and expensive. This paper presents a deep learning model and a methodology for solving such complex classification tasks with only few labelled examples. We show in particular that mixing one-shot learning with recurrent neural networks and an attention mechanism enables obtaining efficient models while preserving some form of interpretability and limiting potential overfit. Results obtained on several types of claims in French court decisions, using different vectorization processes, are presented.