An Improved Transformer-Based Neural Machine Translation Strategy: Interacting-Head Attention

Figure 2

Comparison between multihead attention (a) and our mechanism (b) with 4 heads.

(a)

(b)