The article is actually quite forthright. Singhal implies that the core of this uses some form of machine learning. Looks like a binary good-bad classifier that

The article is actually quite forthright. Singhal implies that the core of this uses some form of machine learning. Looks like a binary good-bad classifier that can give confidence or a probability in its predictions. Negative ratings have a higher threshold. I'm thinking what is more important than classifier is what features are chosen. I bet the breakthrough was a clever way to choose or identify these features.