Scaling Up Semantic Structure for NLP

The UCI Center for Machine Learning is proud to present Julian Michael, Ph.D. student, University of Washington. UCI campus community is invited to join in person in DBH 6011. Everyone else is invited to join Zoom: https://uci.zoom.us/j/94313342769

Title:
Scaling Up Semantic Structure for NLP

Abstract:

Formal representations of linguistic structure and meaning can provide basic building blocks for characterizing language behavior and building robust, explainable models for NLP tasks — in theory. In practice, such representations have proven difficult to specify with broad coverage, annotate reliably, and use in downstream tasks. In this talk, I will describe recent efforts to address these challenges using an approach called Question-Answer driven Semantic Role Labeling (QA-SRL). I will cover two key advantages of QA-SRL over traditional semantic representations. First, it denotes semantic relations using natural language question-answer pairs, which can be annotated by non-experts and applied directly downstream. Second, it imposes carefully-chosen grammatical constraints on its questions which make them highly engineerable, allowing for large-scale annotation, controllable generation of expressive natural language questions, and even the automated induction of formal ontologies of semantic roles. Reflecting on this work, I will discuss the lessons it provides for task design and data annotation, and future directions for the development of scalable semantic representations.

Bio:

Julian Michael is a PhD student at the University of Washington, where he primarily works on new approaches for representing semantic structure. He obtained his BS in computer science from the University of Texas at Austin, and has worked in research labs at Meta, Google, the Allen Institute for AI, and Adobe. His research interests are broad, including work on explainable modeling, benchmarking, model analysis, question answering, and formal logic, and he has a general interest in the philosophy of science as it applies to AI and NLP. Outside of research, you can usually find him rock climbing, skiing, or reading about medicine and psychology.

https://julianmichael.org/