We analyze Recurrent Neural Network (RNN) architectures to handle the problem of Part-of-Speech (POS) Tagging. When linguistic rules are inserted ad-hoc into the decision algorithm, there is a difficulty in understanding the role of prior information and learning. The real potential of recurrent networks is demonstrated in this paper on the Italian language in a purely data-driven approach, where we can reach the state-of-the-art on the UD_Italian-ISTD (Italian Stanford Dependency Treebank) dataset in comparison to TINT. We propose a methodology for splitting words that are mapped to embedding spaces and fed to forward-backward networks.