Skip to Main Content
Table 3 
BLEU scores for multilingual Europarl models with various sizes k of the attention bridge. For comparison, the table includes results of a multi-way multilingual NMT model (Firat, Cho, and Bengio 2016) and a completely shared architecture with language labels: LL (Johnson et al. 2017).
  M ↔ ENM ↔ EN
k = 1k = 10k = 25k = 50FiratLLk = 1k = 50LL
  DE 14.66 19.87 20.61 20.83 18.49 21.63 14.89 20.47 21.7 
EN ES 21.82 27.55 28.41 28.13 27.73 29.48 21.4 27.6 29.53 
  FR 17.8 23.35 24.36 23.79 23.22 25.56 17.62 24.15 25.51 
  
DE   16.97 21.39 23.42 24 24.8 25.96 17.38 24.4 25.84 
ES EN 18.38 25.39 27.01 27.12 25.7 28.41 19.43 26.98 28.67 
FR   17.52 21.93 24.4 23.9 24.52 26.93 17.47 24.47 25.47 
  M ↔ ENM ↔ EN
k = 1k = 10k = 25k = 50FiratLLk = 1k = 50LL
  DE 14.66 19.87 20.61 20.83 18.49 21.63 14.89 20.47 21.7 
EN ES 21.82 27.55 28.41 28.13 27.73 29.48 21.4 27.6 29.53 
  FR 17.8 23.35 24.36 23.79 23.22 25.56 17.62 24.15 25.51 
  
DE   16.97 21.39 23.42 24 24.8 25.96 17.38 24.4 25.84 
ES EN 18.38 25.39 27.01 27.12 25.7 28.41 19.43 26.98 28.67 
FR   17.52 21.93 24.4 23.9 24.52 26.93 17.47 24.47 25.47 
Close Modal

or Create an Account

Close Modal
Close Modal