Linguistic seminar series September 2018
- Date: –17:00
- Location: Engelska parken 9-2029
- Lecturer: Spencer Caplan
- Contact person: Marc Tang
The Acquisition of Vowel Harmony from Simple Local Statistics (and why you might care)
Abstract: Vowel harmony denotes a class of phonotactic constraints which limit which vowels can co-occur in words. The characteristics of harmony systems have been well-researched from theoretical, typological, and developmental perspectives (Bakovic, 2000; Kaun, 2004, inter alia.). Children are sensitive to harmony very early in their development, as young as seven months (Mintz et al., 2018), so the mechanisms responsible for harmony acquisition must be able to identify its presence as well as the specifics of individual vowel harmony systems with little input. Prior computational work (Baker, 2009) has sought either to detect the presence of harmony without describing the specific implementation or to describe a specific implementation when the general details are known beforehand. In this talk I present a new computational psycholinguistic acquisition model inspired by phonological notions of restrictiveness (Caplan & Kodner 2018). Taking in only a stream of unsegmented phonemes as input, the model succeeds in automatically detecting harmony and describing the gross characteristics of the underlying harmony grammar without prior knowledge about the type of system to expect. This explicit harmony acquisition model has impact outside of solely the language acquisition domain as it independently derives several attested typological tendencies associated with harmony languages (Aksenova & Deshmuk, 2018) which I will also discuss.