This dissertation investigates the mechanisms that determine perspective in human sentence processing. It is claimed that processing perspective involves three different kinds of mechanism: a pragmatic and processing default that presumptively attributes attitudes to the speaker of the utterance, a set of partially conventionalized surface cues that signal that the speaker intends to contravene this default, and a general abductive inferencing procedure that selects the best representation of the general context from what is known about the perspectives of individual agents in the discourse, in combination with what was asserted by the speaker. To this end, I develop the notion of an 'agent profile' which stores known and anticipated information about individuals in memory, one of which determines a perspectival parameter in a general, non-linguistic representation of context. I show that attributing perspective exclusively to a non-speaker agent may be difficult to achieve, but once achieved, is the most economical decision in certain environments. Experimental results from a wide range of expressions and constructions are presented, including epithets, appositives, and matrix reports which convey a non-speaker perspective, as in free indirect discourse.