Combining shallow and deep learning approaches against data scarcity in legal domains