This Friday, Peter Graff will give a practice talk for his upcoming CLS paper (with Florian Jaeger). Please note the special time and location!
Speaker: Peter Graff (with Florian Jaeger)
Title: The OCP is a pressure to keep words perceptually distinct: Evidence from Javanese
Time: Friday 4/10 3:30pm, 32-D831
In this study we advance two claims about co-occurrence restrictions on consonants (OCP; Leben 1973) based on a case study of Javanese: i) belonging to the same perceptually salient natural class significantly decreases the likelihood of two consonants co-occurring, ii) that this probabilistic penalty increases linearly with the number of similar segments within a root evidencing cumulativity of OCP effects. Generalizing from perceptual experiments, we hypothesize that the OCP functions as a lexical optimization constraint to keep the words of a language perceptually distinct.
In the first part of this study we investigate whether perceptually salient natural classes have stronger OCP effects associated with them than other sets. In order to not over-parameterize the model we chose a subset of possible natural classes, some with perceptual correlates (e.g. rhotic, lateral, strident) and some with articulatory correlates (e.g. alveolar, glide, palatal). Of 9,261 theoretically possible C1VC2VC3-templates, 1,913 are attested (Uhlenbeck, 1978). We use logistic regression to test whether C1VC2VC3-templates where any two of C1, C2, C3 belong to a natural class are less likely to occur. We simultaneously control for the frequency of C1, C2, and C3 in their respective positions as well as identity (C1=C2), which is known to be favored in Javanese. We find highly significant OCP effects of both articulatory and perceptually motivated classes. By far the strongest similarity avoidance effects, however, are observed for features that are independently known to be highly perceptually salient (rhotic-/r/ and lateral-/l/, Heid and Hawkins 2000; β/r/=-3.86,p<0.0001; β/l/=-2.47,p<0.0001; cf. mean β’s for other OCP effects=-1.47).
Gallagher (2008) shows that, for some features, listeners are better at discriminating words with 0 instances of a feature from words with 1 or 2, than at distinguishing words with 1 instance from words with 2. . Given this result, we generalize that if the OCP is a pressure to optimize perceptual distinctness of words, then additional similar segmentsmake roots even less likely. Indeed, model comparison shows that a cumulative model explains the data significantly better than a non-cumulative model (Bayesian Information Criterion difference=82.5).
Our aim is to place this study in a larger context of logistic regression models of five more languages on which we are currently conducting similar studies. We hope to see i) whether perceptually salient classes of segments co-occur less and ii) whether OCP effects are cumulative as expected under our hypothesis. We will compare our models to other models of similarity avoidance (Frisch et al. 2004, Coetzee and Pater 2008) to see whether our generalizations hold up independent of modeling approach and whether any of these models has an inherent advantage in predicting possible roots.