Interactive property attribution in concept combination

Abstract

We address the question of how people understand attributive noun-noun compounds. Alignment-and-comparison models suggest that the similarity of the constituent concepts guides interpretation. We propose, as an alternative, an interactive property attribution model wherein the modifier and head concepts have different functions: The head provides relevant dimensions, whereas the modifier provides candidate features for attribution. According to our model, the interaction of dimensions and features, rather than constituent similarity, guides interpretation. In this study, we empirically contrasted the two models by holding constituent similarity of compounds constant while varying the interaction of modifier feature salience and head dimension relevance. Compounds consisting of a head concept with a relevant dimension for attribution and a modifier with a salient property on that dimension were interpreted by means of property attribution. Other compounds with equivalent constituent similarity, but lacking the high salience-relevance interaction, were not interpreted by means of attribution. The interactive property attribution model more accurately predicted interpretation of noun-noun compounds.