Sentiment Neuron Linear model + L1 Lp Regularization Uses very few learned units Single sentiment neuron that predicts the sentiment value Can be useful for Unsupervised Learning Refs