Author here. I wanted to explore the GBooks ngram dataset, and see what parts of sentences were less predictable by models, so I made a little visualization.