Table 5 Component-wise ablation study of the proposed model on CommonsenseQA.

From: Knowledge-based question answering using graph neural networks and contextual language representations

Model variant

Accuracy (%)

 ± std (3 seeds)

Full model (GATv2 + pruning + projection fusion, joint fine-tuning)

82.3

 ± 0.3

Remove GNN (text-only BERT classifier)

78.2

 ± 0.4

Replace GATv2 with vanilla GAT

81.1

 ± 0.35

Full model without pruning (unpruned subgraphs)

81.6

 ± 0.35

Full model, GAT frozen (no joint fine-tuning)

80.7

 ± 0.45

Full model, naive concatenation (no projection)

81.9

 ± 0.3

Back to article page