Dependency-aware self-attention for robust neural machine translation.

The results, summarized in Table 4 , reveal a clear trend: model performance improves as DASA is introduced into progressively higher layers, peaking at the third layer.

also in 19,460 other papers