Verbs like start semantically select for event complements like the meeting; if these verbs are not followed by events, such as in the verb phrase (VP) started the newspaper, readers must coerce the complement into events using enriched composition. Some forms of enriched composition can be primarily based on local (VP-internal) semantic properties (local coercion), while others must rely on more global (VP-external) semantic properties (discourse coercion), e.g., properties derived from the subject. If event senses are primarily based on local properties of the NP complement and if local semantic properties are retrieved more automatically, then local coercions may be easier to process than event senses that are based on more global, discourse properties. This study used a self-paced reading time procedure to examine whether processing of the two forms of enriched composition differed. The results indicate that when both VP internal and external properties were available, expressions involving discourse coercion did not yield longer reading times than those involving local coercion (Experiment 1); however, when external properties were not available, discourse coercion was more difficult than local coercion (Experiment 2). The results suggest that coercion based on local properties is no easier than coercion based on external properties, but local information provides a default interpretation when discourse constraints are weak or unavailable.



Dave Roofeh (Advisor: Bob Rehder)

Research on categorization has shown effects from both empirical and theoretical knowledge on predicting unobserved features in category exemplars. Recent cognitive models suggest that theoretical causal knowledge involves Bayesian networks, in which people reason on the basis of causal relations to predict the presence or absence of unobserved category features. In contrast, empirical or similarity-based feature induction holds that the strength of the inferences depends on the similarity of the exemplar to the prototype, presenting a non-normative violation of the causal Markov condition stipulated by the Bayesian-view of causal knowledge. It is hypothesized that if the similarity-to-prototype model mediates feature prediction for object categories, non-object categories should produce results more inline with the Bayesian-view of causal knowledge. Participants predict features of and classify three non-object categories (societies, economies, and weather systems); half receive causal knowledge about the categories.

This study examines the Bayesian view of causal reasoning for feature prediction and will prompt further research on 1) the nature of the relation of non-object categories to current theories of knowledge or 2) other possible models to account for feature prediction.