Joint Source-Target Self Attention with Locality Constraints

Benchmark Model Rank Results
machine-translation-on-iwslt2014-germanLocal Joint Self-attention#18BLEU score: 35.7
machine-translation-on-wmt2014-english-frenchLocal Joint Self-attention#10BLEU score: 43.3
machine-translation-on-wmt2014-english-germanLocal Joint Self-attention#19BLEU score: 29.7