The texts are first represented by 10k TFIDF word scores that represent word affinity.
These features are then feeded in a standard dense neural network, with 6 layers and gradually fewer neurons. Finally, the network outputs a single chance for being a positive review, between 0 and 1.
This notebook contains this pipeline from preprocessing movie reviews to building a sentiment classifier model. It classifies 88% validation reviews correctly - but does not do equally well on short reviews.