Epistemic adjectives: Lexical semantics meets the psychology of reasoning
Epistemic adjectives – those expressing (un)certainty, such as 'possible', 'likely', 'certain' – are vague and admit of degree modification, like many non-modal adjectives. Recently a number of authors have observed that their modification behavior and the inferences that they license problematize received theories of epistemic modality in a number of ways (e.g., Yalcin 2010; Lassiter 2010, 2014). These authors have conclude that (at least some) epistemic adjectives should be treated as degree expressions which place conditions on the probability of their argument, analogously to the scalar adjectives 'heavy' and 'full', which place conditions on the weight and fullness (resp.) of the object that they are applied to.
Independently, the adjectives 'plausible', 'necessary', and 'certain' have been used in reasoning experiments in an effort to show that ordinary reasoning behavior involves two qualitatively different kinds of reasoning, one 'analytic' and one 'associative' (Heit & Rotello 2010; Rips 2001; Rotello & Heit 2009). The motivating observation is that perceived argument strength, as estimated by rate of endorsement, depends in a non-linear way on whether participants are instructed to judge arguments by whether their conclusions are 'plausible' given the premises, or whether they are 'necessary'.
In the first half of this talk I motivate the probabilistic solution to the lexical semantic problem using data involving degree modification and inferential behavior. In the second half (based on joint work with Noah Goodman) I show that this semantics, combines with a probabilistic treatment of vagueness to predict the possibility of non-linearities while maintaining a one-process, Bayesian theory of reasoning. This theory makes a strong monotonicity prediction which is not shared by two-process theories, and I present a novel reasoning experiment which verifies this prediction. Rather than problematizing one-process theories of reasoning generally, participants' behavior when reasoning with probability modals appears to provide direct support for a specific one-process theory, based on Bayesian inference.