Artificial Intelligence

T1.Constraints for Transfer Learning for Machine Translation T2. Say Yes-and: Building a Specialized Corpus for Digital Improvised Comedy

Friday, August 24, 2018, 3:00pm - 4:00pm PDTiCal
NL Seminar
This event is open to the public.
NL Seminar
Mozhdeh Gheini, Xinyu Wang (ISI intern)


Abstract: Can we detect the parts responsible for a generic behavior in a model to transfer it to another? In this talk, we first see why this might be a good idea, especially for low-resource machine translation. Then we focus on our approach to isolating a behavior. In our case, we specifically focus on coverage during machine translation. We present our results across different languages that show how neural models try to ensure coverage.

Bio: Mozhdeh Gheini is a last-semester Computer Science master's student at USC Viterbi School of Engineering. At ISI, she works on improving neural low-resource machine translation under the supervision of Jonathan May. She will be applying for Ph.D. programs this Fall.


Abstract: In improvised comedy, saying "yes, and.. " is a rule-of-thumb that suggests that one person should accept the other person's offer (yes), and then add related information on top of that (and). Collecting a "yes, and.." corpus is not only helpful for building an improv agent, but can also be used for building conversational skill training tool, improving a dialogue system, etc. I will discuss the methods we have used for building such a dataset, data we have got so far and future considerations.

Bio: Xinyu is a 2018 summer intern working with Dr. Jonathan May and Dr. Nanyun Peng on computerized improvised comedy. She will be joinging the Language Technologies Institute at Carnegie Mellon University in 2018 fall.


« Return to Events