Machine Learning / Duolingo Seminar
Speaker
WILLIAM W. COHEN
Principal Scientist
Google Research
Consulting Professor, Machine Learning Department/Language Technologies Institute
Carnegie Mellon University
When
-
Where
In Person and Virtual Presentation - ET
Description
The goal of open-domain question answering (QA) is simply to answer questions about any subject. Although very large language models like GPT-3 or PaLM can answer many questions directly, using information stored only in their parameters, the best-performing methods on most datasets use an “open-book” approach, where documents are retrieved from a large text corpus and then read and reasoned with to produce the answer. Systems that retrieve from a KB also exist, but tend to have poor performance on most QA benchmarks, because many questions are not well-aligned with the contents of existing KBs. A recent alternative approach to QA is to retrieve from a collection of previously-generated question-answer pairs. Question-answer pairs (QA pairs) are appealing in that they seem to be an intermediate between text and KB triples: like KB triples, they usually concisely express a single relationship, but like text, they have good coverage.
I will describe two active lines of work that build on this idea. First, I will describe how a collection of QA pairs can be extended to function like an open-domain, textually grounded, relational KB. Second, I will discuss in depth one of the key technical problems associated with QA-pair-based QA systems, namely the problem of learning to retrieve, which is a latent step in the QA process, and present some alternative approaches to learning this retrieval step. I will also describe initial results in answering compositionally-defined “multi-hop” questions, which do not explicitly appear in the collection of stored QA pairs.
This is joint work with Wenhu Chen, Nitish Gupta, Pat Verga, John Wieting, and others.
—
William Cohen received his bachelor's degree in Computer Science from Duke University in 1984, and a PhD in Computer Science from Rutgers University in 1990. From 1990 to 2000 Dr. Cohen worked at AT&T Bell Labs and later AT&T Labs-Research, and from April 2000 to May 2002 Dr. Cohen worked at Whizbang Labs, a company specializing in extracting information from the web. From 2002 to 2018, Dr. Cohen worked at Carnegie Mellon University in the Machine Learning Department, with a joint appointment in the Language Technology Institute. Dr. Cohen is a past president of the International Machine Learning Society, was program chair or general chair of the International Machine Learning Conference 3 times, is a AAAI Fellow, and winner of two “Test of Time” awards, one from SIGMOD and one from SIGIR. Dr. Cohen is now a Principal Scientist at Google, and is a Consulting Professor at the School of Computer Science at Carnegie Mellon University.
In Person and Zoom Participation. See announcement.