This repository contains the implementation of Text-Relational Graph Neural Networks (Text-RGNN) under review in IEEE SPL for text classification tasks. The model leverages heterogeneous graph neural networks to capture complex relationships in text data, significantly improving performance on benchmark datasets.
Dataset | Split | 1% | 5% | 10% | 20% | 100% |
---|---|---|---|---|---|---|
cola | train | 32.33 | 44.79 | 53.71 | 63.62 | 70.15 |
val | 26.49 | 47.22 | 51.80 | 63.17 | 69.66 | |
test | 38.14 | 47.73 | 56.31 | 61.94 | 68.30 | |
mr | train | 85.65 | 87.16 | 88.04 | 90.71 | 92.38 |
val | 83.58 | 86.49 | 87.43 | 89.96 | 91.62 | |
test | 83.91 | 86.41 | 87.51 | 88.35 | 89.98 | |
ohsumed | train | 68.90 | 77.99 | 82.28 | 85.28 | 92.28 |
val | 59.32 | 70.54 | 71.08 | 75.00 | 81.16 | |
test | 49.45 | 65.52 | 63.29 | 67.33 | 72.86 | |
R8 | train | 97.79 | 98.16 | 97.05 | 97.70 | 98.80 |
val | 97.13 | 98.83 | 97.78 | 96.61 | 97.70 | |
test | 96.48 | 97.81 | 97.44 | 97.76 | 98.86 | |
R52 | train | 94.57 | 97.20 | 97.06 | 97.38 | 98.82 |
val | 91.32 | 96.92 | 96.70 | 94.62 | 96.02 | |
test | 87.46 | 93.89 | 95.06 | 95.44 | 96.85 | |
SST2 | train | 88.60 | 91.09 | 92.78 | 93.57 | 95.45 |
val | 88.77 | 91.38 | 92.89 | 93.49 | 95.37 | |
test | 90.60 | 91.74 | 93.69 | 94.38 | 96.28 |