IWSLT 2016 Machine Translation

Preprocesssing

Moses tokenizer
Moses truecaser
Moses clean-corpus-n
BPE splitting using subword-nmt
- number of merge operations: 16000

Model Configuration

Model: Transformer
- number of layers: 6
- d_model (adim): 256
- d_ff (eunits/dunits): 2048
- Gradient Clipping: 5
- Dropout Rate: 0.1
- Warmup steps: 8000

Result

These results will be updated soon.

		En-->De			De-->en
Framework	tst2012	tst2013	tst2014	tst2012	tst2013	tst2014
Fairseq	27.73	29.45	25.14	32.25	34.23	29.49
ESPNet	26.92	28.88	24.70	32.19	33.46	29.22

Pretrained Model