Past Event

Learning and Generalizing Complex Phonotactics in Artificial Languages

Dr. Kevin McMullin
(University of Ottawa)

Recent studies have demonstrated a relationship between the typology of long-distance phonotactic dependencies, and the way humans learn those patterns in the lab. For example, experiments on sibilant harmony (Finley 2011, 2012, McMullin and Hansson 2014b) as well as liquid harmony and dissimilation (McMullin 2016) show that learners generalize nonlocal patterns of (dis)agreement in CVCVCV contexts into relatively local CVC items, but not vice versa—a result that mirrors an observed typological asymmetry (Rose and Walker 2004, Hansson 2010, Bennett 2013). However, it is unclear how best to treat long-distance dependencies within phonological theory. The Agreement by Correspondence (Rose and Walker 2004) framework, implemented in Optimality Theory, does an excellent job of accounting for the attested patterns of consonant harmony (Hansson 2010) and long-distance consonant dissimilation (Bennett 2013), but predicts a number of pathological systems that are unattested and computationally complex (McMullin and Hansson in press, McMullin 2016). Alternative approaches that treat these patterns as members of the Tier-based Strictly Local class of formal languages (Heinz et al. 2011) seem to improve typological predictions while offering an explanation of learnability, but omit the possibility of certain patterns that have (limited) cross-linguistic support. Sundanese, for example, which Bennett (2014) analyzes with a surface correspondence approach, exhibits a pattern of liquid harmony in LVL contexts, but liquid dissimilation in LVCVLV contexts (Cohn 1992). The goal of this study is to investigate these limits on inductive phonotactic learning by presenting subjects with complex patterns (including a Sundanese-like pattern) in the lab, in order to better understand the empirical data that phonological theory needs to address. In the present experiment, participants are tasked with learning ‘verb conjugations’ in an artificial language. The ‘past’ and ‘future’ tense of C1VC2VC3V roots are formed by adding the suffixes [-li] and [-ru], respectively. These suffixes trigger liquid alternations in the root, resulting in a pattern of relatively local liquid harmony between C3 and Csuff, but a nonlocal pattern of liquid dissimilation that holds between C2 and Csuff. The training phase, in which subjects hear and repeat verb triplets, thus includes items like {depile; depile-li, depiru-ru} and {bilono; birono-li, bilono-ru}. Subjects are then tested to determine whether they prefer liquid harmony or dissimilation at each level of locality in the verb root (i.e. CVCVLV-LV, CVLVCV-LV, LVCVCVLV). Results thus far indicate that subjects who learn the target dissimilation pattern in the nonlocal CVLVCV-LV testing items also apply the target harmony pattern in the local CVCVLVLV contexts. Preliminary results suggest that the same is true for subjects exposed to nonlocal harmony and local dissimilation. These findings will be compared to those of McMullin 2016, which suggest that learners should over-generalize and apply the same pattern to all levels of locality (in spite of the patterns observed in training).

About the Speaker

Dr. Kevin McMullin received his PhD in linguistics from UBC in 2015. He works in computational phonology and formal language theory, and has done experimental work on artificial language learning. He has recently started a tenure-track position in linguistics at the University of Ottawa.