67 - GLUE: A Multi-Task Benchmark and Analysis Platform, with Sam Bowman
NLP Highlights
00:00
Why Leaderboards Encourage Bad Science
I feel like things like Elmo or language models are just inherently more scalable to information content than a single vector. And you're right that for some applications, say I want to have some huge database of sentences that I can do quick lookups in this vector space. But even then, the particular vector that you're going to use is going to have some extracted features that is a fixed length set. And so however you pre-train this is going to be focused on some particular objective, and hopefully it matches how you want to actually use it.
Transcript
Play full episode
