Harnessing the linguistic signal to predict scalar inferences