English compound stress in an analogical model of word formation

Some English noun-noun compounds, e.g. ápple juice and téabag, usually have stress on the lefthand constituent while others, e.g. apple píe and car rádio, normally have stress on the righthand constituent. Recent corpus studies have shown that, inter alia, a compound's semantics and the identities of its constituent nouns are significant probabilistic predictors of stress placement (e.g. Plag et al. 2008; Plag 2010). Furthermore, English compound stress is subject to within-type variation, so that many compounds are attested with both right and left stress patterns, though some show more within-type variation than others (Kunter 2011). We show that these empirical facts of English compound stress are naturally accounted for by an analogical theory of morphophonology, implemented computationally in the analogical algorithm AM::Parallel (Skousen & Stanford 2007). The evidence is based on 1,000 compounds from the British National Corpus, experimentally elicited from multiple speakers and carefully rated for stress placement. In a series of simulation experiments, AM correctly predicts over 90% of both left and right stresses in the portion of this database that does not show within-type variation. Furthermore, in the dataset as a whole, AM successfully predicts the extent to which different types are variably stressed. An analogical theory assumes that new words are formed on the basis of similar items already in the lexicon. For compound stress, this means that stress is assigned to new compounds in accordance with the stress pattern of similar compounds previously encountered. In the AM model presented here, degree of similarity is calculated using the compounds' constituents and semantic properties; close analysis of the model reveals how these different factors interact in an analogical system. Some compounds with similar semantic properties tend to cluster into a large 'analogical gang', influencing stress assignment in new compounds with the same semantic properties, and overriding potential constituent effects. The result resembles a categorical rule. For many other compounds, however, stress assignment is based on fewer more highly similar analogues: in these cases, the identities of the constituents are central in the computation of similarity. Based on these findings, it is argued that an analogical theory provides a grammatical model in which the major types of effect seen in English compound stress assignment are expected. These include within-type variability, specific effects of particular constituents (e.g. compounds ending in 'cake' are usually left stressed), and more general semantic effects that resemble abstract rules. In an analogical model, all of these effects arise naturally from a single mechanism.