1. We will use an encoder (either a CNN or an RNN) to map each string of text (hypothesis and premise) to a fixed-dimension vector representation.

2. We will interact the two hidden representations and output a 3-class soft- max. (To keep things simple, we will simply concatenate the two repre- sentations, and feed them through a network of 2 fully-connected layers.)

Data: https://nlp.stanford.edu/projects/snli/