Bayesian modeling of spatial typology

• Applicant: Dr. Matías Guzmán Naranjo
• Subject Area: General and Comparative Linguistics, Experimental Linguistics, Typology, Non-European Languages
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 504155622

The role of contact and areal effects on shaping languages has long been recognized in the dialectological and typological literature. In typology in particular, the existence of Sprachbunds and contact zones has been of great interest for the past 40 or so years, and there is a considerable amount of research on linguistic areas for all continents. While studies on language contact and areal effects cover different perspectives, there are, in broad terms, two foci of study with respect to these issues: (1) controlling for contact in the search of universals, and (2) and exploring areal features themselves. The first focus is related to the question of sampling and independence. When searching for language universals and relative distributions of linguistic feature, typologists need to control for family effects and contact effects. If in a typological sample two languages have been in contact, we cannot assume that they are independent observations. For this reason, statistical models are often used to try to control for the spatial non-independence of languages. The second focus is not about trying to control for contact when investigating universals, but rather it is about exploring how contact between languages and geographic features leads to sprachbunds, areal patterns, and feature diffusion in space. In this case the objective is to automatically find contact patterns from the data. Both foci have been extensively explored in typology with the use of different statistical tools and models. However, attempts at building computational models of these issues have remained disconnected, and for the most part have relied on faulty or incomplete assumptions, and heavily simplified data. First, most approaches to these issue assume that languages are single points in space, but this assumption is evidently not correct. Second, statistical models of spatial issues often ignore geographic features like pathways and barriers, or fail to build models which realistically capture the interactions between societies and geography. In order to better understand spatial phenomena we need to take these issues seriously and develop more realistic, generative models for spatial typology using more complete and detailed dataset. The present project aims to improve the state of the art by building Bayesian Generative models of spatial phenomena (both for induction and as typological controls) using more realistic assumptions and better quality spatial data. We will work on both modeling to control for spatial-induced confounds, and on understanding spatial effects onto themselves.

• DFG Programme: Independent Junior Research Groups