Multilingual language models, or MLMs, are machine learning models that can predict, generate, and extract text from more than one language. As language models grow larger, their performance improves—as long as they only operate in a single language. "Giving each child a full set of paints to express themselves—or perform tasks in their language—would require massive amounts of pigment, or model parameters. Crucially, this allows for a significant reduction in a language model's size without compromising its performance. Other co-authors of this work include Benjamin Van Durme, an assistant professor of computer science and a member of HLTCOE and CLSP; his advisee Yunmo Chen, a third-year PhD student in Computer Science; Weiting Tan, a doctoral candidate in Computer Science; and Shuyue Stella Li, Engr '22, '23 (MS), a former research assistant at CLSP.
