Alex's NMT Contributions

Experiments:

WMT14 en-fr:

• cell size = 1000, embedding size = 620
• Vocab = 30k source and target
• Max length = 50 source and target
• CGRU: Coditional GRU (Sennrich et al 2017)

IWSLT14 de-en:

• Vocab = 30k source and target
• Max length = 45 source(de) and 47 target(en)

References:

1. Ranzato's settings: Encoder = Convolutional attentive encoder (Rush et al.)

2. Actor-Critic (Bahdanau et al 2017) Data:

• LL = log-likelihood training
• AC = Actor-Critic
• RF-C = Reinforce-Critic

With a cnn encoder (similar to Ranzato)

With a bi-GRU encoder (dim=256) (Wiseman & rush 2016)