Machine Translation on IWSLT (Multi-lingual Pairs)
35.4BLEU (de-en)GHT
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GHTParam=44M, Inference Speed=1016.4 sent/s, FLOPs=1996M2023.05 | 35.4 | — | — | — | — | — | — | — | — | 32.8 | 29.1 | 31.6 | 41.5 | |
| GHT-PSParam=30M, Inference Speed=1122.1 sent/s, FLOPs=1558M2023.05 | 35.2 | — | — | — | — | — | — | — | — | 32.7 | 28.9 | 31.6 | 41.4 | |
| Dynamic convParam=30M, Inference Speed=1050.2 sent/s, FLOPs=1615M2023.05 | 34.8 | — | — | — | — | — | — | — | — | 32.7 | 28.7 | 31.1 | 40.6 | |
| Li et al. (2018)Param=44M, Inference Speed=1011.2 sent/s, FLOPs=1996M2023.05 | 34.7 | — | — | — | — | — | — | — | — | 31.8 | 28.5 | 30.7 | 40.7 | |
| Cordonnier et al. (2020)Param=44M, Inference Speed=416.6 sent/s, FLOPs=2054M2023.05 | 34.4 | — | — | — | — | — | — | — | — | 31.8 | 28.2 | 31 | 40.7 | |
| Lite TransformerParam=30M, Inference Speed=1096.6 sent/s, FLOPs=1809M2023.05 | 33.3 | — | — | — | — | — | — | — | — | 31.4 | 27.5 | 29.8 | 39.4 | |
| Li et al. (2021)Param=30M, Inference Speed=1116.9 sent/s, FLOPs=1558M2023.05 | 33.2 | — | — | — | — | — | — | — | — | 31.3 | 27.5 | 30 | 39.7 | |
| Voita et al. (2019)Param=30M, Inference Speed=1099.1 sent/s, FLOPs=1558M2023.05 | 32.2 | — | — | — | — | — | — | — | — | 30.8 | 26.5 | 30.3 | 39.8 | |
| BaselineModel Architecture=Transformer-small2021.05 | — | 16.9 | 16.4 | 20.9 | 29 | 30.9 | 28.1 | 29.2 | 35.2 | — | — | — | — | |
| LaSSModel Architecture=Transformer-small2021.05 | — | 17.9 | 17 | 22.9 | 30.9 | 33 | 29.8 | 30.9 | 37.3 | — | — | — | — |