Discover and read the best of Twitter Threads about #ACL2020
Most recents (1)
Is your NLP classifier actually (un)biased? Or is your diagnosis based on too little data?
It might be the latter!
In my #ACL2020 paper, I discuss why we need bigger datasets for conclusively identifying classification bias in NLP.
arxiv.org/abs/2004.12332 1/
It might be the latter!
In my #ACL2020 paper, I discuss why we need bigger datasets for conclusively identifying classification bias in NLP.
arxiv.org/abs/2004.12332 1/
Background: Large NLP datasets don't come with annotations for protected attributes (e.g., gender). To test for classification bias, one typically annotates a small sample of data (typically < 5K). WinoBias and WinoGender are great examples of these bias-specific datasets. 2/
Intuitively, the less data we annotate, the less certain we are that our estimate is close to the true bias. But how can we quantify this uncertainty? 3/
