As a possible improvement which may (I have not researched this at all yet) be novel, we could look into funkier ways of representing the questions than simple LSTM+word vectors. The ideas that came into my mind (again, no idea if they have been done yet or not; I can survey the literature move thoroughly over the next few days) are:
Using Facebook's fasttext library to compute representations for words and feed them into an LSTM, OR to directly compute sentence representations (the "Bag of Tricks" paper on text classification had nearly state-of-the-art results in text classification, but required only a fraction of the training time).
Using character-level LSTMs (or GRUs) to process the questions.
As a possible improvement which may (I have not researched this at all yet) be novel, we could look into funkier ways of representing the questions than simple LSTM+word vectors. The ideas that came into my mind (again, no idea if they have been done yet or not; I can survey the literature move thoroughly over the next few days) are:
What do you guys think?