I'm really over the hype around RNNs for NLP tasks - they're just a fancy way of saying "we don't have a clue how to model temporal dependencies, so let's throw some LSTMs at the problem