Logical Complexity in Morphological Learning

1. Introduction Language learning is to a large extent learning how to classify objects based on their properties (features). For instance, learning a morphological paradigm can be viewed as learning conditions for affix insertion, where the distribution of each affix is determined by morpho-syntactic features. Sometimes the distribution of an affix can be described simply in terms of a single conjunction of features (e.g., uses in [3rd person singular] contexts), but sometimes, due to syncretism, an affix has a difficult to state, heterogeneous distribution. A natural hypothesis is that paradigms with simpler affix distributions should be acquired faster, with less errors, and, therefore, be less prone to historical change. However, what is the relevant metric of simplicity (or complexity) for the human learners? We address this question by way of artificial grammar learning experiments that provide a controlled setting for studying what factors affect pattern complexity. Similar types of experiments have a long history in psychology. In particular, there is an extensive literature on learning of artificial categories defined by visual features such as shape, color, size, and so on. The robust findings in this literature can serve as a starting point for identifying what linguistic patterns are more complex than others. Namely, we can test whether the results found for non-linguistic patterns extend to the linguistic domain. This line of inquiry not only lets us investigate complexity of linguistic * I am grateful to Elliott Moreton for many insights related to this work.