Table 2 CORAL accuracy across weak supervision coverage. Training with more weak supervision significantly improves performances
From: CORAL: COde RepresentAtion learning with weakly-supervised transformers for analyzing data analysis
Model | Weak Supervision Coverage | ||
|---|---|---|---|
25% | 50% | 100% | |
CORAL | 41.6 | 46.4 | 72.2 |
CORAL(No Masked Attention) | 31.7 | 46.1 | 70.4 |
BERT(AST, No Finetuning) | 26.1 | 47.2 | 67.9 |