Comment by textminer

If using raw bag of words or n-grams, why not hash the strings to maybe 2^13-1 slots, with something like MurmurHash3 or with multiple hashes to prevent collisions, then use that sparse vector as input to a deep learning model?