T. O’Donnell, Kenny Smith

Evolang

Paper

Abstract

While inflectional morphology is broadly rule-governed, many inflectional paradigms admit some exceptions (e.g., the past tense of “go” is not “goed” but “went”). Regularity in form-meaning mapping permits generalization and facilitates learning, and regularity has been shown to emerge through language transmission (e.g. Kirby et al., 2008, 2015). The presence of irregulars is more puzzling. One possibility is that irregulars are simply a by-product of processes which destroy regularity (e.g., minimisation of effort in production), which survive language transmission due to their high frequency (Kirby, 2001; Lieberman et al., 2007). Here we show that irregularity in the right place in a morphological system actually facilitates the learning of the productive parts of the system; therefore, irregularity might be favoured during language transmission, rather than simply being a by-product of other processes. We build on a recent computational model (O’Donnell, 2015) which shows that high-frequency irregulars facilitate the learning of productive regular rules. This model treats learning as an inference problem, where learners infer the productivity of morphological processes, balancing a tradeoff between computation and storage. High-frequency items tend to be stored as wholes, rather than handled compositionally; productive computation is signaled by morphological processes which apply across a large number of low-frequency forms. Since there is pressure to store high-frequency regular forms, they detract from the productivity of regular inflectional processes. When high-frequency forms are instead irregular, regular rules generalize more easily. We provide experimental evidence for this irregularization bias. 46 adult participants learned 48 novel inflected words, organized into six disjoint paradigms of eight words each (see Table 1). In two Fully Regular sets, all stems were inflected with a single regular suffix and occurred with equal frequency during training. The remaining four sets had eight stems that occurred with non-uniform frequency. In each set, seven stems occurred with a regular suffix while a single item took an irregular suffix. These sets differed in whether the the most frequent 352