Modelling semantic transparency in English compound nouns

Semantic transparency is known to play an important role in the storage and processing of complex words (e.g. Marslen-Wilson et al. 1994), and human raters of transparency achieve high levels of agreement (e.g. Frisson et al. 2008, Munro et al. 2010). In the case of noun-noun compounds, overall transparency is largely determined by the transparency of the individual constituents. For example, Reddy et al. (2011) showed that the perceived transparency of a compound is highly correlated with both the sum and the product of the perceived transparencies of its constituents. Furthermore, many psycholinguistic studies find significant effects for semantic transparency using a four-way distinction based on perceived constituent transparency: transparent-transparent (e.g. carwash), transparent-opaque (e.g. jailbird), opaque-transparent (e.g. strawberry) and opaque-opaque (e.g. hogwash) (Libben et al. 2003). Bell and Schäfer (2013) modelled the transparency of individual compound constituents and showed that shifted word senses reduce perceived transparency, while certain semantic relations between constituents increase it. However, this finding is problematic in at least two ways. Firstly, it is not clear whether there is a solid basis for establishing whether a specific word sense is shifted or not. For example, card in credit card is clearly shifted if viewed etymologically, but may not synchronically be perceived as shifted due to its frequent use. Secondly, work on conceptual combination by Gagné and collaborators has shown that relational information in compounds is accessed via the concepts associated with individual modifiers and heads, rather than independently of them (e.g. Spalding et al. 2010 for an overview). This leads to the hypothesis that it is not whether a specific word sense is etymologically shifted, nor whether a specific semantic relation is used per se, that makes a compound constituent more or less transparent; rather, it is the degree of expectedness of a particular word sense and a particular relation for a given constituent. In this paper, we provide evidence in support of this hypothesis: the more expected the word sense and relation for a constituent, the more transparent it is perceived to be.