In episode 65 of The Gradient Podcast, Daniel Bashir speaks to Sewon Min.
Sewon is a fifth-year PhD student in the NLP group at the University of Washington, advised by Hannaneh Hajishirzi and Luke Zettlemoyer. She is a part-time visiting researcher at Meta AI and a recipient of the JP Morgan PhD Fellowship. She has previously spent time at Google Research and Salesforce research.
Have suggestions for future podcast guests (or other feedback)? Let us know here!
Subscribe to The Gradient Podcast: Apple Podcasts | Spotify | Pocket Casts | RSS
Follow The Gradient on Twitter
Outline:
(00:00) Intro
(03:00) Origin Story
(04:20) Evolution of Sewon’s interests, question-answering and practical NLP
(07:00) Methodology concerns about benchmarks
(07:30) Multi-hop reading comprehension
(09:30) Do multi-hop QA benchmarks actually measure multi-hop reasoning?
(12:00) How models can “cheat” multi-hop benchmarks
(13:15) Explicit compositionality
(16:05) Commonsense reasoning and background information
(17:30) On constructing good benchmarks
(18:40) AmbigQA and ambiguity
(22:20) Types of ambiguity
(24:20) Practical possibilities for models that can handle ambiguity
(25:45) FaVIQ and fact-checking benchmarks
(28:45) External knowledge
(29:45) Fact verification and “complete understanding of evidence”
(31:30) Do models do what we expect/intuit in reading comprehension?
(34:40) Applications for fact-checking systems
(36:40) Intro to in-context learning (ICL)
(38:55) Example of an ICL demonstration
(40:45) Rethinking the Role of Demonstrations and what matters for successful ICL
(43:00) Evidence for a Bayesian inference perspective on ICL
(45:00) ICL + gradient descent and what it means to “learn”
(47:00) MetaICL and efficient ICL
(49:30) Distance between tasks and MetaICL task transfer
(53:00) Compositional tasks for language models, compositional generalization
(55:00) The number and diversity of meta-training tasks
(58:30) MetaICL and Bayesian inference
(1:00:30) Z-ICL: Zero-Shot In-Context Learning with Pseudo-Demonstrations
(1:02:00) The copying effect
(1:03:30) Copying effect for non-identical examples
(1:06:00) More thoughts on ICL
(1:08:00) Understanding Chain-of-Thought Prompting
(1:11:30) Bayes strikes again
(1:12:30) Intro to Sewon’s text retrieval research
(1:15:30) Dense Passage Retrieval (DPR)
(1:18:40) Similarity in QA and retrieval
(1:20:00) Improvements for DPR
(1:21:50) Nonparametric Masked Language Modeling (NPM)
(1:24:30) Difficulties in training NPM and solutions
(1:26:45) Follow-on work
(1:29:00) Important fundamental limitations of language models
(1:31:30) Sewon’s experience doing a PhD
(1:34:00) Research challenges suited for academics
(1:35:00) Joys and difficulties of the PhD
(1:36:30) Sewon’s advice for aspiring PhDs
(1:38:30) Incentives in academia, production of knowledge
(1:41:50) Outro
Links:
Papers
Solving and re-thinking benchmarks
Language Modeling
Text representation/retrieval
Sewon Min: The Science of Natural Language