University of Massachusetts Amherst

Search Google Appliance

Links

Mohit Iyyer Joins CICS Faculty

Mohit Iyyer joined the College of Information and Computer Sciences in September of this year. Prior to coming to UMass, Mohit was a Young Investigator at AI2, and received his PhD at the University of Maryland, College Park. His interests lie broadly in natural language processing, and using machine learning to generate language at the discourse level.
While much of the existing work on language generation is focused on single sentences, Iyyer is interested in generating text at the paragraph or document scale. This idea introduces linguistic issues -- in a four-sentence paragraph, the sentences need to be consistent with each other, and not introduce new concepts or contradict past sentences. In addition, most current models take an inordinate amount of time to train, and conventional wisdom says that the models need to be bigger and consume more data. But is that true? Iyyer and a student are working on a project showing that, "If we train the model on just 1% of the data, but we carefully select that 1% to include things like complex syntactic structures or lots of different types of words, it will result in the same performance."
When asked what he enjoys about his work, Iyyer said he enjoys meeting with students to brainstorm and go over results, determine why a model isn't working correctly, and what to try next. In his spare time he plays a lot of basketball and enjoys playing R&B music on his guitar. Visit his website to learn more about his work.