A Surprisingly Robust Trick for Winograd Schema Challenge

Open paper
Benchmark Model Rank Results
coreference-resolution-on-winograd-schemaBERTwiki 340M (fine-tuned on WSCR)#26Accuracy: 72.5
coreference-resolution-on-winograd-schemaBERT-large 340M (fine-tuned on WSCR)#27Accuracy: 71.4
coreference-resolution-on-winograd-schemaBERTwiki 340M (fine-tuned on half of WSCR)#29Accuracy: 70.3
coreference-resolution-on-winograd-schemaBERT-base 110M (fine-tuned on WSCR)#42Accuracy: 62.3
natural-language-inference-on-wnliBERTwiki 340M (fine-tuned on WSCR)#11Accuracy: 74.7
natural-language-inference-on-wnliBERT-large 340M (fine-tuned on WSCR)#13Accuracy: 71.9
natural-language-inference-on-wnliBERT-base 110M (fine-tuned on WSCR)#14Accuracy: 70.5

Rank counts only results with a code link.