Attention Is (not) All You Need for Commonsense Reasoning

Benchmark Model Rank Results
coreference-resolution-on-winograd-schemaBERT-base 110M + MAS#47Accuracy: 60.3
coreference-resolution-on-winograd-schemaUSSM + Supervised DeepNet + KB#57Accuracy: 52.8
coreference-resolution-on-winograd-schemaUSSM + KB#59Accuracy: 52
natural-language-understanding-on-pdp60BERT-base 110M + MAS#3Accuracy: 68.3
natural-language-understanding-on-pdp60USSM + Supervised Deepnet + 3 Knowledge Bases#4Accuracy: 66.7
natural-language-understanding-on-pdp60USSM + Supervised Deepnet#8Accuracy: 53.3