Questions for Self-monitoring: RNN and LSTM

• What's the difference between recurrent and recursive NN?
• What are the benefits of RNN compared to window-based feedforward NNs?
• What is the most frequently used RNN architecture for NLP and what are its characteristics?
• How training of and inference in simple RNN is performed?
• To which kinds of NLP tasks RNN can be applied?
• How do the training data for these tasks look like?
• How does a stacked RNN architecture look like?
• What's the benefit of a stacked RNN?
• How does a bidirectional RNN look like?

• What are the major disadvantages of simple RNN?
• [Couldn't the problem be solved by using more nodes on the hidden layer?]
• How does the LSTM address these problems?