Grouped Bert for Multi-label Classification
SemEval 2023 Task 4: ValueEval: Identification of Human Values behind Arguments
Instead of the modelling the task as a classification problem, we modeled as Natural Language Inference (NIL) problem which determines the logical relationship (human values- output labels) between the argument and the conclusion.
Designed Grouped-BERT for Multi-label identification of values behind arguments.
Results
| Model | Precision | Recall | F1 | Accuracy |
|---|---|---|---|---|
| 1-Baseline | 0.18 | 1.0 | 0.28 | 0.18 |
| SVM | 0.30 | 0.30 | 0.3 | 0.77 |
| BERT | 0.39 | 0.30 | 0.34 | 0.84 |
| L-label classifier | 0.29 | 0.48 | 0.36 | 0.76 |
| L-Binary classifiers | 0.3 | 0.45 | 0.35 | 0.77 |
| Hybrid + CE loss | 0.32 | 0.42 | 0.39 | 0.77 |
| Hybrid + CE + HD loss | 0.33 | 0.43 | 0.40 | 0.77 |
