GLUE development set results for different architectures for incorporating semantic information. The settings and metrics are identical to Table 3a. All models use the base size variant.
. | MNLI . | . | ||||||||
---|---|---|---|---|---|---|---|---|---|---|
Models . | CoLA . | MRPC . | RTE . | SST-2 . | STS-B . | QNLI . | QQP . | ID. . | OOD. . | Avg. . |
RoBERTa | 63.1 | 90.1 | 79.0 | 94.6 | 91.0 | 93.0 | 91.8 | 87.7 | 87.3 | 86.4 |
GCN | 65.2 | 90.2 | 80.2 | 94.8 | 91.1 | 92.9 | 91.8 | 87.8 | 87.7 | 86.8 |
GAT | 63.4 | 90.0 | 79.4 | 94.7 | 91.2 | 92.9 | 91.8 | 87.7 | 87.6 | 86.5 |
Hidden | 64.2 | 90.2 | 79.7 | 94.5 | 91.0 | 92.8 | 91.8 | 87.1 | 86.7 | 86.4 |
Scaffold | 62.5 | 90.5 | 71.1 | 94.3 | 91.0 | 92.6 | 91.7 | 87.7 | 87.6 | 85.5 |
SIFT | 64.8 | 90.5 | 81.0 | 95.1 | 91.3 | 93.2 | 91.9 | 87.9 | 87.7 | 87.0 |
SIFT-Light | 64.1 | 90.3 | 80.6 | 94.7 | 91.2 | 92.8 | 91.7 | 87.7 | 87.6 | 86.7 |
. | MNLI . | . | ||||||||
---|---|---|---|---|---|---|---|---|---|---|
Models . | CoLA . | MRPC . | RTE . | SST-2 . | STS-B . | QNLI . | QQP . | ID. . | OOD. . | Avg. . |
RoBERTa | 63.1 | 90.1 | 79.0 | 94.6 | 91.0 | 93.0 | 91.8 | 87.7 | 87.3 | 86.4 |
GCN | 65.2 | 90.2 | 80.2 | 94.8 | 91.1 | 92.9 | 91.8 | 87.8 | 87.7 | 86.8 |
GAT | 63.4 | 90.0 | 79.4 | 94.7 | 91.2 | 92.9 | 91.8 | 87.7 | 87.6 | 86.5 |
Hidden | 64.2 | 90.2 | 79.7 | 94.5 | 91.0 | 92.8 | 91.8 | 87.1 | 86.7 | 86.4 |
Scaffold | 62.5 | 90.5 | 71.1 | 94.3 | 91.0 | 92.6 | 91.7 | 87.7 | 87.6 | 85.5 |
SIFT | 64.8 | 90.5 | 81.0 | 95.1 | 91.3 | 93.2 | 91.9 | 87.9 | 87.7 | 87.0 |
SIFT-Light | 64.1 | 90.3 | 80.6 | 94.7 | 91.2 | 92.8 | 91.7 | 87.7 | 87.6 | 86.7 |