Natural language is replete with nuanced evocative language such as word-play, irony, metaphor, and hyperbole. In all of these tropes the conveyed meaning diverges from the overt meaning to convey layers of affective and social information. Yet our best wide-coverage NLP systems either ignore these aspect of meaning entirely, or treat them in a cursory manner. In this talk I will show that detailed probabilistic models of language understanding can form the substrate for quantitative models of humor (puns), hyperbole, and metaphor. Each of these models achieves good quantitative accuracy at predicting human behavioral data. They rely on measurements of non-linguistic background knowledge and I will argue that properly collecting and using these non-linguistic priors will be crucial for making NLP systems able to understand and use human evocative language.