Data Validation specifications are mainly made of sentences where verbal groups give some constraints to be verified. For the purpose of an automated treatment of these natural language specifications, we need to extract and identify those constraints using natural language processing tools. We present in this article an experimentation with a neural network model based on BERT fine-tuning. A list of constraints to identify and a corpus of sentences and syntactic propositions have been created, and a paraphrase generator has been used to balance the lack of training data. Results are promising but can be nevertheless improved, for example by increasing the quantity of data.