CompSci Colloquium: Alvin Grissom II on "Opening Up Linguistic Machine Learning Models"
Opening Up Linguistic Machine Learning Models: Some Difficulties and Rewards
ABSTRACT: In this talk, I will present some of my recent published research, focusing particularly on two projects: the first examines the use of machine learning to incrementally predict verbs in verb-final languages, such as Japanese, in which the main verb of the sentence appears at the end of sentences. Previous research has shown that human simultaneous interpreters do this when translating from verb-final (subject-object-verb) languages to verb-medial (subject-verb-object) languages, such as English, and doing so is useful for machines to learn simultaneous interpretation.
The second project focuses in examining pathologies in deep neural network models for natural language which make them difficult to interpret. Recently, deep neural networks have largely supplanted linear models for many tasks in natural language processing, but the large number of parameters and complexity of the architecture make understanding their behavior extremely difficult. We uncover examples of pathological behavior in such models which exacerbate these interpretative difficulties and offer a simple method of mitigating them for several natural language tasks.
Finally, I will briefly discuss other work on Japanese word segmentation, sentence rewriting for simultaneous translation, digital liberal arts, as well as discuss future research directions.
BIO: Alvin Grissom II is a computational linguist and Assistant Professor of Computer Science at Ursinus College, a small liberal arts college. He does research at the intersection of computational linguistics and machine learning. He received his Ph.D. in computer science from the University of Colorado Boulder in 2017.
Tuesday, January 29 at 3:30pm to 4:30pm
Muenzinger Psychology, MUEN D430
1905 Colorado Avenue, Boulder, CO 80309