Past Events

Qualifying Exam

Can Document Statistics Disambiguate Grounded Language: The Case of PP attachment


Download as iCal file

Thursday, May 27, 2021, 10:00am - 11:30am


Speaker: John Blackmore

Location : Remote via Webex


Prof. Matthew Stone (advisor)

Prof. Martin Farach-Colton

Prof. Karl Santos

Prof. Yongfeng Zhang

Event Type: Qualifying Exam

Abstract: Advancements in the coverage and efficacy of pre-trained language models have paved the way for future research in natural language processing by making it much easier generalize from small hand-annotated datasets. In cases that require inputs from situational context, however, pre-trained models may not reliably learn to disambiguate grounded language. Using the case of PP attachment in image captions, I assess the learning potential of pre-trained statistical language models toward resolving ambiguity in image captions. My results show that pre-trained models may lack the context required to resolve the ambiguity in many cases, but traditional NLP approaches can be combined with statistical models for grounded learning.