We explore the implications of a particular approach to learning for an architectural question in phonology. The learning approach follows the principle of Minimum Description Length (MDL), which has recently been used for learning in both constraint-based and rule-based phonology. The architectural question on which we focus is whether the grammar allows language-specific statements to be made at the level of the lexicon, as was assumed in early generative phonology, or whether such statements are prohibited, as is commonly assumed within more recent work. We show that under MDL, the architectural question has real empirical implications: across a range of seemingly natural representational schemes, an ability to make language-specific statements about the lexicon is needed to ensure the learnability of an important aspect of phonological knowledge.

This paper presents a new criterion based on prediction error which allows the estimation of the number of parameters as well as structures in statistical models. The criterion is valid for short and long samples alike. Unlike Akaike's earlier criterion, also based on prediction error, the criterion proposed here appears to produce consistent error estimates in ARMA processes.