List your company
ResearchFormerly WizWand

Machine Translation on WMT16 English-German (test)

41.2BLEUSOTA (Supervised)
4.43613.980523.52533.0695Apr 24, 2018Dec 20, 2018Aug 17, 2019Apr 14, 2020Dec 10, 2020Aug 7, 2021Apr 5, 2022
Updated 8mo ago

Evaluation results

MethodLinks
SOTA (Supervised)Training=Supervised2020.0541.2
Finetuned SOTAprotocol=Supervised2022.0441.2
PaLM 540Bshots=52022.0437.4
SEQUENCEM=6, N=12, #Params=61M, Speed=x1.312021.0434.99
CYCLE (REV)M=6, N=18, #Params=61M, Speed=x0.982021.0434.84
CYCLE (REV)M=6, N=12, #Params=61M, Speed=x1.312021.0434.81
SEQUENCEM=6, N=18, #Params=61M, Speed=x0.982021.0434.8
CYCLEM=6, N=18, #Params=61M, Speed=x0.982021.0434.79
CYCLEM=6, N=12, #Params=61M, Speed=x1.312021.0434.77
UniversalM=1, N=6, #Params=63M, Speed=x1.002021.0434.6
VanillaM=6, N=6, #Params=61M, Speed=x2.022021.0434.23
Vanilla (deep)M=18, N=18, #Params=149M, Speed=x0.962021.0434.19
Universal (deep)M=1, N=12, #Params=63M, Speed=x0.522021.0434.01
Vanilla (big)M=6, N=6, #Params=210M, Speed=x0.812021.0433.4
Universal (small)M=1, N=6, #Params=24M, Speed=x2.522021.0432.81
Supervised NMTUSMT Tuning=Supervised, Sentence pairs=5.6M2018.1032.3
PaLM 540Bshots=02022.0431.8
PaLM 540Bshots=12022.0431.8
Supervised NMTUSMT Tuning=Supervised, Sentence pairs=2.8M2018.1031.6
Supervised NMTUSMT Tuning=Supervised, Sentence pairs=1.4M2018.1029.9
mBARTTraining=Unsupervised2020.0529.8
GPT-3Few-shot learning=Few-shot2020.0529.7
MASSTraining=Unsupervised2020.0528.3
XLMEncoder Pre-training=MLM, Decoder Pre-training=-2019.0127
FLANshots=112022.0426.8
XLMEncoder Pre-training=MLM, Decoder Pre-training=MLM2019.0126.4
XLMTraining=Unsupervised2020.0526.4
GPT-3Few-shot learning=One-shot2020.0526.2
GPT-3 175Bshots=12022.0426.2
FLANshots=02022.0425.4
XLMEncoder Pre-training=MLM, Decoder Pre-training=CLM2019.0124.9
GPT-3Few-shot learning=Zero-shot2020.0524.6
XLMEncoder Pre-training=-, Decoder Pre-training=CLM2019.0124.4
Supervised trainingTraining mode=Supervised, Loss=Standard cross-entropy2018.0424.07
XLMEncoder Pre-training=CLM, Decoder Pre-training=CLM2019.0122.7
UNMT (this work)USMT Tuning=Supervised, Filtering=Enabled2018.1021.6
XLMEncoder Pre-training=MLM, Decoder Pre-training=-2019.0121.6
UNMT (this work) w/o filteringUSMT Tuning=Supervised2018.1021.3
XLMEncoder Pre-training=EMB, Decoder Pre-training=EMB2019.0121.3
XLMEncoder Pre-training=CLM, Decoder Pre-training=MLM2019.0121.3
Lample et al. (2018b) USMT+UNMTUSMT Tuning=No2018.1020.2
PBSMT + NMT2019.0120.2
UNMT (this work)USMT Tuning=No, Filtering=Enabled2018.1020
UNMT (this work) w/o filteringUSMT Tuning=No2018.1019.6
XLMEncoder Pre-training=CLM, Decoder Pre-training=-2019.0119.2
Artetxe et al. (2018b) USMTUSMT Tuning=back-translation2018.1018.2
Lample et al. (2018b) USMTUSMT Tuning=No2018.1017.8
PBSMT2019.0117.8
USMT (this work) w/ forward translationUSMT Tuning=Supervised2018.1017.4
Lample et al. (2018b) UNMTUSMT Tuning=No2018.1017.2
NMT2019.0117.2
USMT (this work) w/ back-translationUSMT Tuning=Supervised2018.1017
USMT (this work) w/ forward translationUSMT Tuning=No2018.1015.5
USMT (this work) w/ back-translationUSMT Tuning=No2018.1015
Unsupervised Neural Machine Translation with Weight SharingWeight-sharing layers=12018.0410.86
Lample et al.2018.049.64
XLMEncoder Pre-training=-, Decoder Pre-training=-2019.016.7
Word-by-wordMethod type=Baseline2018.045.85