Skip to Main Content
Table 3 
Word translation accuracy (P@1) of Arabic and low-resource (Ms, He) languages on the Conneau data set using FastText embeddings. ** indicates failure to converge.
 En-ArEn-MsEn-He
Supervised Baselines 
Artetxe, Labaka, and Agirre (2017) 24.8 45.3 38.8 41.6 32.7 52.1 
Artetxe, Labaka, and Agirre (2018a) 41.2 55.2 55.1 51.7 47.6 58.0 
Procrustes-CSLS 34.5 49.7 47.3 46.6 39.2 54.1 
  
Unsupervised Baselines 
Hoshen and Wolf (2018) 34.4 49.3 ** ** 36.5 52.3 
Artetxe, Labaka, and Agirre (2018b) 33.2 52.8 49.0 49.7 43.8 57.5 
Conneau et al. (2018) (code) 29.3 47.6 46.2 ** 36.8 53.1 
  
Our Unsupervised Approach 
Adversarial autoencoder + 
 Conneau refinement 33.8 49.9 49.5 48.6 41.1 56.8 
 Artetxe refinement 38.3 54.1 54.0 54.4 44.9 58.1 
 Our combined refinement 38.6 55.7 54.8 55.2 46.1 58.6 
 En-ArEn-MsEn-He
Supervised Baselines 
Artetxe, Labaka, and Agirre (2017) 24.8 45.3 38.8 41.6 32.7 52.1 
Artetxe, Labaka, and Agirre (2018a) 41.2 55.2 55.1 51.7 47.6 58.0 
Procrustes-CSLS 34.5 49.7 47.3 46.6 39.2 54.1 
  
Unsupervised Baselines 
Hoshen and Wolf (2018) 34.4 49.3 ** ** 36.5 52.3 
Artetxe, Labaka, and Agirre (2018b) 33.2 52.8 49.0 49.7 43.8 57.5 
Conneau et al. (2018) (code) 29.3 47.6 46.2 ** 36.8 53.1 
  
Our Unsupervised Approach 
Adversarial autoencoder + 
 Conneau refinement 33.8 49.9 49.5 48.6 41.1 56.8 
 Artetxe refinement 38.3 54.1 54.0 54.4 44.9 58.1 
 Our combined refinement 38.6 55.7 54.8 55.2 46.1 58.6 
Close Modal

or Create an Account

Close Modal
Close Modal