| Aodong LI |
Only make the English encoder's embedding constant -- 45.98
Only initialize the English encoder's embedding and then finetune it -- 46.06
Share the attention mechanism and then directly add them -- 46.20
- big data baseline bleu = 30.83
- Model with three fixed embeddings
Shrink output vocab from 30000 to 20000 and best result is 31.53
Train the model with 40 batch size and best result until now is 30.63
|
- test more checkpoints on model trained with batch = 40
- train model with reverse output
|
| Shiyue Zhang |
- trained word2vec on big data, and directly used it on NMT, but resulted in quite poor performance
- trained M-NMT model, got bleu=36.58 (+1.34 than NMT). But found the EOS in mem has a big influence on result:
| NMT
|
35.24, 57.7/39.8/31.9/27.0 BP=0.939
|
| MNMT (EOS=1)
|
35.27, 60.0/41.3/33.1/28.0 BP=0.907
|
| MNMT (EOS=0.2)
|
36.40, 59.1/40.8/32.6/27.4 BP=0.951
|
| MNMT (EOS=0)
|
36.58, 58.4/40.4/32.1/27.0 BP=0.968
|
- tried to tackle UNK using 36.58 M-NMT, increased vocab to 50000, got bleu=35.63, 58.6/40.0/31.6/26.4 BP=0.953 (not good, ?)
- training uy-zh, 50% zh-uy, 25% zh-uy
- training mem without EOS
- reviewing related papers
|
|
