A Deep Generative Model of Vowel Formant Typology

Proceedings of NAACL-HLT 2018, pages 37–46 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics A Deep Generative Model of Vowel Formant Typology Ryan Cotterell and Jason Eisner Department of Computer Science Johns Hopkins University, Baltimore MD, 21218 {ryan.cotterell,eisner}@jhu.edu Abstract What makes some types of languages more probable than others? For instance, we know that almost all spoken languages contain the vowel phoneme /i/; why should that be? The field of linguistic typology seeks to answer these questions and, thereby, divine the mech- anisms that underlie human language. In our work, we tackle the problem of vowel system typology, i.e., we propose a generative proba- bility model of which vowels a language con- tains. In contrast to previous work, we work di- rectly with the acoustic information—the first two formant values—rather than modeling dis- crete sets of phonemic symbols (IPA). We de- velop a novel generative probability model and report results based on a corpus of 233 lan- guages. 1 Introduction Human languages are far from arbitrary; cross- linguistically, they exhibit surprising similarity in many respects and many properties appear to be universally true. The field of linguistic typology seeks to investigate, describe and quantify the axes along which languages vary. One facet of language that has been the subject of heavy investigation is the nature of vowel inventories, i.e., which vowels a language contains. It is a cross-linguistic univer- sal that all spoken languages have vowels (Gordon, 2016), and the underlying principles guiding vowel selection are understood: vowels must be both easily recognizable and well-dispersed (Schwartz et al., 2005). In this work, we offer a more formal treatment of the subject, deriving a generative prob- ability model of vowel inventory typology. Our work builds on (Cotterell and Eisner, 2017) by in- vestigating not just discrete IPA inventories but the cross-linguistic variation in acoustic formants. The philosophy behind our approach is that lin- guistic typology should be treated probabilistically and its goal should be the construction of a univer- sal prior over potential languages. A probabilistic approach does not rule out linguistic systems com- pletely (as long as one’s theoretical formalism can describe them at all), but it can position phenomena on a scale from very common to very improbable. Probabilistic modeling also provides a discipline for drawing conclusions from sparse data. While we know of over 7000 human languages, we have some sort of linguistic analysis for only 2300 of them (Comrie et al., 2013), and the dataset used in this paper (Becker-Kristal, 2010) provides simple vowel data for fewer than 250 languages. Formants are the resonant frequencies of the hu- man vocal tract during the production of speech sounds. We propose a Bayesian generative model of vowel inventories, where each language’s inven- tory is a finite subset of acoustic vowels represented as points (F1, F2) ∈ R2. We deploy tools from the neural-network and point-process literatures and experiment on a dataset with 233 distinct languages. We show that our most complicated model outper- forms …

Content truncated to 500 words