Thu, Dec 09, 2021 @ 11:00 AM - 01:00 PM
PhD Candidate: Yury Zemlyanskiy
Title: Parametric and semi-parametric methods for knowledge acquisition from text
Time: 11:00 AM-1:00 PM PST, Dec 9 (Thursday)
Committee: Fei Sha, Leana Golubchik, Xiang Ren, Robin Jia, Jonathan May, and Meisam Razaviyayn (external).
Zoom link: https://usc.zoom.us/j/95672050503
Knowledge acquisition is a crucial characteristic of an intelligent system that allows the processing of large amounts of information. Nonetheless, modern neural networks (e.g., BERT) used in natural language processing typically do not have a dedicated memory component. The knowledge about the world that the models acquire is stored implicitly in the model's parameters. This proves unreliable and makes the models ill-suited for knowledge-intensive tasks that require reasoning over vast amounts of textual data.
My thesis explores alternative parametric and semi-parametric methods to extract and represent knowledge from text. Specifically, the proposed methods seek to establish several desirable properties for the neural network memory component. First, the memory should benefit the model and allow it to reason over large amounts of textual information. Second, the memory should be amendable and adapt to a new context (a different book or collection of articles) on the fly. Finally, certain applications will benefit from the transparent structure of the memory, allowing queries on information about particular objects or entities.
The proposed thesis consists of three sections: the first section focuses on parametric memory for a pre-defined set of entities. The second section explores a semi-parametric approach to capturing entity-centric facts in a long document or entire corpus. Finally, the last section discusses future work on memory specialized for structure prediction tasks.
WebCast Link: https://usc.zoom.us/j/95672050503
Audiences: Everyone Is Invited
Contact: Lizsl De Leon