Speaker: Adam Albright
Title: Discovering and modeling cumulative markedness interactions with loglinear models
Date/Time: Monday Mar 19, 5:30pm
Location: 32-D461 (Note unusual location)
It is often observed that “phonology can’t count”. This principle rules out, among other things, languages in which a marked structure is tolerated once or twice within a word, but not three or more times. In this talk, I discuss a set of restrictions in Lakhota (Siouan) which are very similar to such a ‘threshold’ effect: roots often contain a single marked structure (fricative, aspirated or ejective stop, consonant cluster), but roots containing multiple marked structures are rarer than one would expect, based on the independent frequencies of those structures. This observation leads two questions: is the degree of underattestation significant, and if so how should it be accommodated in a grammatical model? I show that both questions can be addressed using log-linear (maximum entropy) models of constraint interaction. First, I present results of a series of statistical models of the Lakhota lexicon, attempting to predict the relative type frequency of root shapes based on their phonological properties. The results show that models with interaction terms, in which multiple simultaneous violations may be penalized more than expected based on the individual violations, do significantly better at predicting lexical counts. Furthermore, the effect is strongest for combinations of structures that are independently most strongly penalized. Thus, it appears that cumulative effects are real, and some form of ‘counting’ is indeed warranted. I argue that these statistical models are too powerful, however: they could, in principle, impose strong penalties on combinations that are independently penalized only weakly (or not at all), or they could even reverse the direction of the preference so that languages tolerate a marked structure only in the presence of another marked structure. I argue that we can avoid these predictions with a simpler model, in which markedness constraints interact with MParse (Prince and Smolensky 1993/2004). I present the results of a learning simulation, showing that the observed cumulative effects can be predicted using a small set of markedness constraints on simple structures. Finally, I consider some typological predictions of the proposed model.