Evaluating and Inducing Dialectal Robustness in Large Language Models

This seminar talk by Niyati Bafna from the Center for Language & Speech Processing (CLSP) at Johns Hopkins University explores how large language models perform across dialects and language variants. Learn about the challenges these models face when handling low-resource dialects despite performing well on high-resource languages. Discover a framework for understanding performance degradation as a function of linguistic distance between related languages using artificially generated dialects. The presentation introduces DialUp, a novel method designed to make pretrained machine translation models more robust across dialect continua, even for previously unseen dialects. Examine the factors that influence the success of this approach and explore future research directions in dialectal robustness for language models.