Terra Blevins

Research Abstract:

While language models (or LMs, à la ChatGPT) grow larger and gain new zero-shot capabilities, performance for non-English languages increasingly lags behind. My research examines the ways in which current language models do and don't capture different languages and characterizes how multilingual language models differ from monolingual ones. Examples of my work in this space include quantifying non-English language contamination in English training data (EMNLP, 2022) and analyzing the pretraining dynamics of multilingual models (EMNLP, 2022). Due to the nature of my research, I also build new methods for analyzing modern NLP systems (ACL, 2023). Finally, my current work focuses on using analysis insights from my prior work to build better, more equitable multilingual systems with sparse training of expert language models.

Bio:

Terra Blevins is a final-year PhD student at the University of Washington, advised by Luke Zettlemoyer. Her research focuses on linguistic analysis of computational models of language and multilingual NLP, with the goal of using analysis insights to build better, more equitable multilingual systems. She received the NSF Graduate Research Fellowship for her research and previously worked as a visiting researcher at Facebook AI Research (FAIR). Terra graduated from Columbia University in 2017 with a BA in Computer Science.