This paper presents a theory of how language is understood, and gives some supporting experimental evidence. Its fundamental hypothesis is that discourse is sometimes mentally represented in a form akin to that of perceived or imagined events. Skilled narrators have the power to elicit such representations so that their audiences seem to experience the events rather than merely to read or hear about them. The theory assumes that there are two main stages in comprehension. First, utterances are translated into a mental code that provides a direct linguistic representation of them. This stage concerns the identification of speech sounds, the recognition of words, and the recovery of superficial syntactic structure. Secondly, the linguistic code may be used as part of the basis for the inferential construction of a mental model of the state of affairs that the utterances describe. On some occasions, listeners go no further than the first stage of interpretation. Several lines of research support the two-stage theory. If people construct a mental model of a discourse, then, for example, their memory for its gist is better than if they have failed to do so, but their recall of verbatim detail is poor. If people do not construct a mental model of a discourse but rely solely on the linguistic code, then they tend to remember the overall import of the passage poorly, but they are often able to recall verbatim detail. Such contrasting results were obtained by comparing determinate descriptions with indeterminate ones that could not be accurately represented by a single mental model. The paper presents a number of other phenomena concerning the coherence of discourse that corroborate the theory.