The human processing of natural language occasionally reflects systematicies of the system’s formal organization, often at the level of syntax. Data on the processing of appositive relative clauses (Dillon et al., 2014; Dillon et al., 2017) suggest that the processor may also be sensitive to the pragmatic organization of information within a sentence: independent speech acts embedded within a maximal speech act appear to be isolated from later parsing and decision making. In three experiments, I examine the processing of another structure which may create an embedded speech act, direct discourse speech reports. Despite existing evidence suggesting direct discourse is indeed processed independently from its superordinate context, the experiments reported here find no support for the prediction that it is isolated from downstream processes. Taking this as evidence against the strong hypothesis that subordinate speech acts serve as hard boundaries in processing, I outline possible alternative hypotheses.

Language comprehension requires a complex series of decisions under uncertainty. This is especially obvious when one string may have multiple different interpretations, whether due to lexical ambiguity, or the potential for an inference beyond literal content. This dissertation profiles how the human system for language comprehension times those decisions, specifically when and why it sometimes postpones them. Evidence comes from nine reading experiments in English probing variation across a range of different types of uncertain meaning (homonymy and polysemy, predicate distributivity, scalar implicatures from "some", and causal inferences from discourse coherence) and across two tasks (self-paced-reading and the Maze task of Forster et al., 2009). Diagnosing the presence of decisions by testing for garden-path effects and costs associated with selection against a bias, I highlight two key patterns. First, some decisions, e.g. sense specification for polysemous nouns, are delayed in normal reading, but occur immediately when a rapid decision would be more useful; I conclude that decisions to postpone are flexible and sensitive to a comprehender's goals. Second, possible scalar implicatures and causal inferences rapidly affect comprehension, but do not receive any typical, decisive commitment until much later; I conclude that comprehenders may develop expectations gradiently based on multiple possible interpretations before they make a firm decision. Throughout the dissertation, I explore how these and related facts might be explained as a consequence of the ways humans attempt to rationally allocate cognitive resources under uncertainty.

Studies on the reading of appositive relative clauses (ARCs) have found that ARCs seem to exhibit less influence in later parsing and decision-making than similar constructions (Dillon et al. 2014, 2017), a pattern we call discounting. Existing accounts often link discounting to the status of ARCs as independent segments in systems of pragmatic representation. This would predict discounting for other constructions as well. In this study, we test that prediction by investigating the reading of direct discourse speech reports and causal adjuncts in English. Diagnostics supplied by the theoretical literature show that these constructions contribute the same independent segments as ARCs in two different systems of pragmatic organization: direct discourse reports contribute an independent speech act, and causal adjuncts contribute their own discourse units. Nevertheless, in a series of five experiments, we find no evidence of ARC-like discounting for either. We conclude that discounting should not be linked to either of these pragmatic representations, and discuss the outlook for other approaches to the phenomenon.