5fa1a76
1
2
We therefore propose a novel span-based dynamic convolution to replace these self-attention heads to directly model local dependencies.
