Title: How to find the rabbit in the big(ger) box:
Reasoning about contextual parameters for gradable adjectives under embedding (joint work with Roger Levy (MIT BCS) and Elizabeth Coppock (Boston University))
Speaker: Helena Aparicio (MIT BCS)
Time: Friday, 22nd of March 2-3pm
Abstract:Haddock (1987) noticed that the definite description ‘the rabbit in the hat’ succeeds in referring even in the presence of multiple hats, so long as only one hat contains a rabbit. These complex definites suggest that uniqueness with respect to the NP hat is not required in such embedded contexts, raising the question of what the correct formulation of the uniqueness condition for definite determines is. Generally speaking, two types of solutions have been proposed to this puzzle. The first one postulates a complex semantic representation for definite determiners, where uniqueness can be checked at different points of the semantic representation for either sets of hats or sets of rabbit-containing hats (Bumford 2017). The second type of account proposes that definite descriptions can be evaluated against a sub-portion of the maximally available context (Evans 2005; Frazier 2008; Muhlstein 2015). This pragmatic mechanism ensures that reference resolution is successful, even when the maximal context would violate the uniqueness presupposition of the definite article.
The present work seeks to tease apart these two classes of theories by investigating the interpretive preferences for similarly embedded noun phrases containing a positive or comparative adjective (e.g., ‘the rabbit in the big/ger box’). Experimental results show that embedded positive adjectives exhibit a sensitivity to contextual manipulations that embedded comparatives lack. We derive this sensitivity using a probabilistic computational model of the contextual parameters guiding the interpretation of the embedded NP, and compare it to alternative models that vary in the lexical representations assumed for definite determiners. Our simulation results show that neither of the two proposals under consideration can independently account for all of the observed experimental results. We show that the model that best matches human data is one that combines both a complex uniqueness check (à la Bumford) with pragmatic context coordination.