The interpretation of adjective-noun compounds is crucial to our ability to make inferences in natural language. In formal semantics, adjectives are often placed in a hierarchy that should dictate their behavior when placed in adjective-noun compounds. For example, subsective adjectives like 'red' add new information that restricts the denotation of the head noun (e.g. a 'red car' is a 'car') while non-subsective adjectives like 'pretend' do not pick out a subset of the denotation of the head noun (e.g. a 'pretend car' is not a 'car').
In this talk, I will discuss experimental findings which suggest that inferences involving adjective-noun composition is not as cut-and-dry as is often assumed. I will focus on two main questions: 1) when can a speaker add an adjective without communicating new information (i.e. without restricting the denotation of the noun)? and 2) when, if ever, does adding or removing an adjective produce a contradictory utterance (i.e. when is the denotation of the modified noun disjoint from that of the unmodified noun)? I will discuss our findings in the context of automatic systems for natural language understanding.