Grounding Systematic Syncretism in Learning

Abstract

It is commonly assumed that patterns of syncretism in inflectional paradigms are restricted in some way. In this article, I show how such restrictions can reflect cognitive constraints on language learning. Namely, I construct a learning algorithm that is biased toward certain types of affix distributions in paradigms, thereby rendering them systematic. In developing this algorithm, I rely on the traditional notions of underspecification and blocking, but recast them in terms of learners' biases toward generalization strategies based on cross-situational intersections and default reasoning. This algorithm allows us to test claims about systematicity of syncretism using typological data and language acquisition studies. In the last part of the article, I present a crosslinguistic survey of verbal agreement paradigms that supports the algorithm's predictions.

The rise and fall of the L-shaped morphome: diachronic and experimental studies

“…The rise and fall of the L-shaped morphome patterns of natural vs. unnatural syncretism and formal modeling of the computational primitives that favor one system over another (e.g Bobaljik (2004), Pertsova (2011), Graf (2012), and in the case of the present paper, adopting the tack of creating closely-controlled experimental conditions in which native speakers must choose between one of two potential, non-existing candidates, or rate the well-formedness of unseen inflected forms of wug words.…”

Machine Learning of Inflection

“…For instance, most patterns of syncretism that cannot be captured by underspecification alone, can be captured by underspecification and blocking (Pertsova, 2011). There are also cross-paradigmatic regularities involving syncretism: in many cases, the same cells tend to be syncretic in different paradigms within the same language revealing that something systematic is going on.…”

“…According to the typological data by Baerman and Brown (2005) from the World Atlas of Language Structures, from a sample of 140 languages that mark subject and person agreement on the verbs, 80 (or 57%) have no syncretism or homonymy in any of the paradigms, which guarantees that every phonologically distinct form can be mapped to a single conjunction of features values. Other typological work on person number marking found that in paradigms with syncretism the most common type of syncretism is the one that can be described by underspecification (Cysouw, 2003;Pertsova, 2011), which is also a conjunctive type of syncretism.…”

“…There is a very simple learning strategy that can learn conjunctions from positive examples alone, while disjunction is best learned from negative examples. This strategy involves looking for features shared among all the instances belonging to the same category (Siskind, 1996;Molnar, 2001;Pertsova, 2011). Such a procedure automatically favors conjunctive categories because every positive instance of a conjunctive category includes all the properties that are shared by the members of the category at large (possibly together with some irrelevant properties).…”

“…To learn disjunctive categories, the learner has to do some extra work, predicting that such categories should be harder to learn. The learner described in Pertsova (2011) introduces another bias on top of the conjunctive bias, namely a bias towards concepts that can be succinctly described as complements of conjunctive concepts (making a prediction that "elsewhere" type patterns should be the second easiest to learn after the conjunctive patterns).…”