Wednesday, February 26, 2020
12:00 PM
209 Computer Studies Building
Wenpeng Yin
Salesforce Research
Universal and Trustworthy Natural Language Understanding through Textual Entailment
Abstract:
Much ink has been spilled on a number of fundamental questions regarding the building of an intelligent machine given task-specific annotated data. This has considerably pushed the frontiers of natural language understanding (NLU). However, we are still far away from our dream—universal artificial intelligence, typically because: i) We study and model each NLU problem independently. Human beings, in contrast, can correlate multiple tasks and leverage knowledge learned in the past; ii) Current learning algorithms are overly dependent on (large-scale) human annotations. We lack a clear solution to make machines work in real-world scenarios where annotations are limited (e.g., zero-shot and few-shot problems), and iii) Despite claims in the media for human level solutions to certain problems, we are not ready to trust the predictions of machines, since complex and deep computational models today often lack interpretability towards the results.

In this talk, I will introduce our progresses towards the goal of universal and trustworthy NLU: i) universal NLU—we try to build a single entailment machine to solve diverse NLU problems, especially when rich annotations are unavailable for the target problems; and ii) trustworthy NLU—we rethink the current evaluations of some tasks, then describe how we develop textual entailment to make them more comprehensive and reliable.

Bio:
Wenpeng Yin is currently a research scientist in Salesforce Research, Palo Alto, California. Before, he was a postdoctoral researcher in University of Pennsylvania, working with Prof. Dan Roth in textual entailment and information extraction. Wenpeng received a Ph.D. degree in Oct. 2017 from Prof. Hinrich Schütze's group in LMU Munich, Germany. He had an internship in IBM Watson Research Center in early 2016, and got multiple competitive awards in the past, including WISE2013 “Best Paper”, "Baidu Ph.D. Fellowship" in 2014, "Chinese Government Award for Outstanding Self-financed Ph.D. Students Abroad" in 2016 and “Area Chair Favorites” paper award in COLING2018. He was invited as the Area Chair for NAACL'2019 and ACL'2019.