Transfer Learning for Language Understanding and Generation

When:
Tuesday, March 31, 2020, 11:00 am - 12:00 pm PDTiCal
Where:
VTC Only
This event is open to the public.
Type:
AI Recruitment Seminar-THIS WILL BE A REMOTE PRESENTATION ONLY
Speaker:
Di Jin (MIT)
Video Recording:
https://usc.zoom.us/rec/share/-sV7bPb3x2ZIeIno9kjyf5N_Mt_maaa823VN-6BYxEuCPG79rLXx1f4GJyonQO71
Description:

Abstract: Deep learning models have been increasingly prevailing in various Natural Language Processing (NLP) tasks, and even surpassed human-level performance in some of them. However, the performance of these models would degrade significantly on low-resource data, even worse than conventional shallow models in some cases. In this work, we combat with the curse of data-inefficiency with the help of transfer learning for both language understanding and generation tasks. First, I will introduce MMM, a Multi-stage Multi-task learning framework for the Multi-choice Question Answering (MCQA) task, which brings in around 10% of performance improvement on 5 MCQA low-resource datasets. Second, an iterative back-translation (IBT) schema is proposed to boost the performance of machine translation models on zero-shot domains (with no labeled data) by adapting from the source domain with large-scale labeled data.

Bio: Di Jin is a fifth year PhD student at MIT working with Prof. Peter Szolovits. He works on Natural Language Processing (NLP) and its applications into biomedical and clinical domains. Previous works focused on sequential sentence classification, transfer learning for low-resource data, adversarial attacking and defense, and text editing/rewriting. 

« Return to Upcoming Events