Predicting scalar inferences from “or” to “not both” using neural sentence encoders