Content extraction performance (Macro F1 on the GROTOAP2 dataset) for I-VILA using different BERT model variants. I-VILA can be applied to both standard BERT-based models and layout-aware ones, and consistently improves the classification accuracy.
| Base Model . | Baseline . | Text Line G(𝓛) . | Text Block G(𝓑) . |
|---|---|---|---|
| DistilBERT | 90.52 | 91.14 | 92.12 |
| BERT | 90.78 | 91.65 | 92.31 |
| RoBERTa | 91.64 | 92.04 | 92.52 |
| LayoutLM | 92.34 | 92.37 | 93.38 |
| Base Model . | Baseline . | Text Line G(𝓛) . | Text Block G(𝓑) . |
|---|---|---|---|
| DistilBERT | 90.52 | 91.14 | 92.12 |
| BERT | 90.78 | 91.65 | 92.31 |
| RoBERTa | 91.64 | 92.04 | 92.52 |
| LayoutLM | 92.34 | 92.37 | 93.38 |