Predicting the outcome of a case from a set of factual data is a common goal in legal knowledge discovery. In practice, solving this task is most of the time difficult due to the scarcity of labeled datasets. Additionally, processing long documents often leads to sparse data, which adds another layer of complexity. This paper presents a study focused on the french decisions of the European Court of Human Rights (ECtHR) for which we build various classification tasks. These tasks consist first of all in the prediction of the potential violation of an article of the convention, using extracted facts. A multiclass problem is also created, with the objective of determining whether an article is relevant to plead given some circumstances. We solve these tasks by comparing simple linear models to an attention-based neural network. We also take advantage of a modified partial least squares algorithm that we integrate in the aforementioned models, capable of effectively dealing with classification problems and scale with sparse inputs coming from natural language tasks.