October 18, 2019
Tepper School Professor Receives Facebook AI Research Award
The award provides Zachary Chase Lipton with nearly $80,000 to support his research.
Zachary Chase Lipton, Assistant Professor of Operations Research and Machine Learning, was awarded a Facebook AI Research Award for his project titled “Robustifying NLP by Exploiting Invariances Learned via Human Interaction.”
Lipton submitted the project in response to Facebook’s call for proposals under “Robust Deep Learning for Natural Language Processing.” He credited Language Technologies Institute Ph.D. students Divyansh Kaushik — whom Lipton advises — and Danish Pruthi for making him aware of the opportunity and insisting he apply.
“The proposal lined up beautifully with recent work with Danish on making NLP models robust to adversarially chosen spelling mistakes and a new project underway with Divyansh that uses humans in the loop in a creative way to make models less reliant on certain spurious associations,” Lipton said.
For this award, Facebook Research sought projects that address major challenges with natural language processing functions such as machine translation, question answering, and content classification. Issues such as variations in spelling, style, and vocabulary in the informal language of online conversations or difficulties arising from noise or bias in training data can challenge AI modeling of natural language. With this award, Facebook sought to support research that offers novel approaches to achieve more robust and higher quality models.
As part of the application process, Lipton was required to provide a brief summary of the project. “The proposal outlined our vision for using humans in the loop to discover and characterize new invariances in natural language data, and to augment datasets so as to break certain spurious associations,” he said, such as a system designed to predict the sentiment of movie reviews that erroneously associates the word “horror” with a negative review. “Our preliminary experiments demonstrate an approach that yields models less dependent on this sort of evidence.”
The award provides Lipton with nearly $80,000 to support his research. “One great thing about these faculty research awards is that there’s a lot of flexibility in how to allocate the funding,” he said. “This helps us to nurture this new line of research and get it off the ground, and hopefully eventually have whole lines of research blossom that can eventually win long-term support.”
“This research exemplifies the interdisciplinary nature of our work at the Tepper School, incorporating applications of language technology, machine learning, and data analysis,” said Alan Scheller-Wolf, Richard M. Cyert Professor of Operations Management and Senior Associate Dean of Faculty and Research. “It’s wonderful to have Facebook’s support for this work. I am eager to see the insights that will come out of this line of research.”
Lipton stressed that the students were a primary driver for the success of their proposal. “I’m honored to be the primary investigator of this team and grateful to have this support for my students,” he said.