Abstract

In West Germanic, noun noun compounds are generally ‘left stressed’, i.e. spoken with prominence on the first element. However, Present-day English has two common patterns: while some noun noun combinations (NNs) show the expected prominence on the first constituent, e.g. ‘table lamp, others receive prominence on the second, e.g. silk ‘shirt. In recent corpus studies by Plag et al. (2007) and Plag (2010), the most reliable predictors of prominence in a given NN were found to be the identities of the constituents: for example, ‘table lamp might be assigned prominence by analogy with ‘table mat, ‘table top etc.

In this talk, I will build on Plag’s work by considering why certain constituents should predispose a compound to right or left prominence. Using data from the British National Corpus (BNC) and from a large scale production experiment, I will present a probabilistic model of prominence assignment that has very good predictive accuracy. The model has three main elements: the semantics of the constituents and the compound as a whole, the relative informativeness of the two constituents, and the length of the second constituent. This model, in particular the contribution of informativeness to prominence assignment, lends support to a view of the lexicon as a highly structured network in which both whole compounds and their constituents are stored. In addition, the results can be interpreted as evidence for an accentual theory of compound stress.