Title: Modeling the Social Dynamics Using Language Change

Sandeep Soni

School of Interactive Computing

College of Computing

Georgia Institute of Technology

Date: Wednesday, May 6, 2020

Time: 3:00 pm – 5:00 pm (EST)

Location: https://bluejeans.com/335386194

Committee:

Dr. Jacob Eisenstein (Advisor,  School of Interactive Computing, Georgia Institute of Technology)

Dr. Munmun De Choudhury (School of Interactive Computing, Georgia Institute of Technology)

Dr. Chao Zhang (School of Computational Science and Engineering, Georgia Institute of Technology)

Dr. Rada Mihalcea (Department of Electrical Engineering and Computer Science, University of Michigan)

Abstract:

All natural languages change over time. The changes in language are systematic, often with social motivations, and strongly related to factors such as social influence and individual or group identities. The study of the social dynamics of language change can thus uncover the latent organization and social structure. As more timestamped text becomes available, large-scale computational modeling of language change can offer us ways to answer socially relevant questions about influence and social identities: Who leads? Who follows? Who diverges?

In this thesis, I contribute to the growing body of using computational approaches to diachronic modeling of language change but with a focus on modeling their social dynamics and assessing the importance of influence and identity as underlying factors. I show that such modeling can not just help in validating known sociolinguistic theories about language change at scale, but also in discovering previously unknown links between semantic change and influence.  In a series of computational and data-driven studies, I first provide quantitative evidence in support of the role social influence plays in language change in online networks. By carefully repurposing contemporary predictive models of event cascades to test network hypotheses at scale, I find that influence from strong ties is more important for language change. I then present a new Granger-causal test for detecting social influence in event cascades that is robust both to the presence of homophily as a confound and practically limiting conditions such as missing data and misspecification. I then offer a new way of identifying instances of progressive usages of semantic innovations in documents further going on to show that semantic progressiveness of US court opinions and scientific articles is strongly correlated with their influence. Building off on these studies, I propose a model to induce a leadership network with respect to semantic changes between a set of anti-slavery newspapers from the nineteenth century and analyses to investigate their role in shaping the discourse. Finally, I propose a study to evaluate whether influence and identity are underlying factors that lead to the abandonment of slang in online social media.