Tweeted By @sleepinyourhat

By that measure, MSR's model is somewhat better than T5 or RoBERTa, but it still falls back on stereotypes substantially more often than humans. We know that LMs pick up stereotypes from their training data, and that's not something we can easily counteract. Proceed with caution.

— Prof. Sam Bowman (@sleepinyourhat) December 30, 2020