CNNs and RNNs for Text

Convolutions learn n-grams. Recurrences remember. Both are superseded by attention. Both still matter on constrained hardware.