CAREER: Sociolinguistic Structure Induction

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

One of the principal functions of language is to manage social relationships. Yet much of our current language technology fails to account for how language varies across social contexts, or how it depends on speakers' attitudes and identities. As a consequence, existing language technology is brittle to variation, particularly in non-standard dialects and informal social contexts. This CAREER project is aimed at closing this gap, by building data-driven computational algorithms and representations that are capable of identifying social meaning in large corpora of text and speech. The scope of the research includes local social variables, such as the address terms used to modulate formality in conversation ("Ms", "dude"), as well as group-level dialect differences across large communities. The expected results of the research include: (1) new computational methods for studying language's social function and, correspondingly, new insights; (2) leveraging these insights to enable new computational linguistic applications, and to improve the robustness of existing technology; (3) better engagement between computer science and the social sciences in augmenting the tools that researchers require in order to access large-scale naturally occurring social and linguistic data; and (4) improved education on the analysis and interpretation of such data for undergraduate and graduate students and the public at large. This project uses unsupervised structure induction techniques to identify the latent structures that connect linguistic and social phenomena. Sociolinguistic structures can be mined from a diverse array of data sources, including social media and digitized humanities archives. The project aims at interactions between linguistic and social phenomena on multiple levels of detail: (1) macro-level sociolinguistic structures, such as the intersecting social identities that drive dialect variation; (2) micro-level sociolinguistic structures, such as the array of address terms that modulate formality in dialogue; and (3) dynamic sociolinguistic structures, such as the transmission of new words, orthographies, and syntactic constructions across social media. Identifying these sociolinguistic structures requires new learning algorithms and representations, driven by theoretical insights from the social sciences. The findings of this project are disseminated in venues that target both the computational and social science communities, and the outreach and education components include building new bridges between these communities.