List your company
ResearchFormerly WizWand

Machine Translation on IWSLT German-to-English '14 (test)

37.6BLEU ScoreTransformer Base & Cutoff
20.055224.610129.16533.7199Jun 9, 2016Sep 7, 2017Dec 7, 2018Mar 7, 2020Jun 5, 2021Sep 4, 2022Dec 4, 2023
Updated 8mo ago

Evaluation results

MethodLinks
Transformer Base & CutoffBackbone=Transformer Base, JS loss=true2020.0937.6
Data Diversification2020.0937.2
Transformer Base & CutoffBackbone=Transformer Base, JS loss=false2020.0936.7
MAT+Knee2020.0936.6
Mixed Representations2020.0936.4
MAT2020.0936.2
AGDArchitecture=Transformer2023.1235.94
AdaBeliefArchitecture=Transformer2023.1235.93
AdamWArchitecture=Transformer2023.1235.82
AdaHessianArchitecture=Transformer2023.1235.79
SGD temperatureArchitecture=6L-Encoder / 6L-Decoder, Optimizer=SGD, Scaling=inverse temperature2022.0635.69
T-FixupArchitecture=6L-Encoder / 6L-Decoder2022.0635.59
Adam POST-LNArchitecture=6L-Encoder / 6L-Decoder, Optimizer=Adam2022.0635.39
Adversarial training2020.0935.2
Adam PRE-LNArchitecture=6L-Encoder / 6L-Decoder, Optimizer=Adam2022.0635.1
APO-PrecondModel=Transformer2022.0234.62
AdamWModel=Transformer2022.0234.6
ReZeroArchitecture=6L-Encoder / 6L-Decoder2022.0634.55
Transformer Base2020.0934.4
Transformerembedding_method=FRAGE2018.0933.97
AdamWimplementation=Fairseq2021.0333.95
SGDHessimplementation=Fairseq2021.0333.73
Variational Attentioninference_type=Enum2018.0733.69
AdaHessianimplementation=Fairseq2021.0333.62
Variational Attentioninference_type=Sample2018.0733.3
Marginal Likelihood2018.0733.29
Transformer2018.0933.12
AdamArchitecture=Transformer2023.1232.93
ConvS2S+Risk2018.0932.93
Minimum Risk Training2018.0732.84
SGD res-scaleArchitecture=6L-Encoder / 6L-Decoder, Optimizer=SGD, Residual Scaling=depth-dependent2022.0632.79
Soft Attention2018.0732.77
ConvS2S+SeqNLL2018.0932.68
Dual transfer learning2018.0932.35
DenseNMT-8L-2 (BPE)Decoding Strategy=Beam, Preprocessing=BPE2018.0632.26
DenseNMT-8L-1 (BPE)Decoding Strategy=Beam, Preprocessing=BPE2018.0632.08
SGDmModel=Transformer2022.0231.43
Hard Attentioninference_type=Enum2018.0731.4
SGD POST-LNArchitecture=6L-Encoder / 6L-Decoder, Optimizer=SGD2022.0631.36
DenseNMT-8L-2 (BPE)Decoding Strategy=Greedy, Preprocessing=BPE2018.0630.8
DenseNMT-8L-1 (BPE)Decoding Strategy=Greedy, Preprocessing=BPE2018.0630.5
Hard Attentioninference_type=Sample2018.0730.42
DenseNMT-8L-2 (word)Decoding Strategy=Beam, Preprocessing=word2018.0630.33
NPMT+LM (this paper)Decoding strategy=Beam Search2017.0630.08
Neural PBMT + LM2018.0730.08
Variational Relaxed Attention2018.0730.05
DeepConv2018.0930.04
NPMT (this paper)Decoding strategy=Beam Search2017.0629.92
SGDimplementation=Fairseq2021.0329.75
NPMT+LMDecoding Strategy=Beam2018.0629.16
DenseNMT-8L-2 (word)Decoding Strategy=Greedy, Preprocessing=word2018.0629.11
NPMTDecoding Strategy=Beam2018.0628.96
SGDArchitecture=Transformer2023.1228.57
NPMT (this paper)Decoding strategy=Greedy2017.0628.57
ACDecoding Strategy=Beam2018.0628.53
AC+LL (Bahdanau et al., 2017)Decoding strategy=Beam Search2017.0628.53
Actor-Critic2018.0728.53
Actor-critic2020.0928.5
RF-C+LL (Bahdanau et al., 2017)Decoding strategy=Beam Search2017.0628.3
NPMTDecoding Strategy=Greedy2018.0627.83
RF-C+LL (Bahdanau et al., 2017)Decoding strategy=Greedy2017.0627.7
LL*Decoding strategy=Beam Search2017.0627.61
LL (Bahdanau et al., 2017)Decoding strategy=Beam Search2017.0627.56
ACDecoding Strategy=Greedy2018.0627.49
AC+LL (Bahdanau et al., 2017)Decoding strategy=Greedy2017.0627.49
BSO, SB-ΔKte=5, Ktr=62016.0626.36
Beam Search Optimization2018.0726.36
LL*Decoding strategy=Greedy2017.0626.17
LL (Bahdanau et al., 2017)Decoding strategy=Greedy2017.0625.82
BSO, SB-ΔKte=10, Ktr=62016.0625.48
BSO (Wiseman & Rush, 2016)Decoding strategy=Beam Search2017.0625.48
NPMT-LMSearch Strategy=Beam Search2017.0625.36
NPMTSearch Strategy=Beam Search2017.0625.08
seq2seqKte=52016.0624.03
Word-level CNN w/attn, input feedingGranularity=Word-level, Attention=With, Input feeding=With2016.1124
seq2seqKte=102016.0623.87
LL (Wiseman & Rush, 2016)Decoding strategy=Beam Search2017.0623.87
BSO, SB-ΔKte=1, Ktr=62016.0623.83
BSO (Wiseman & Rush, 2016)Decoding strategy=Greedy2017.0623.83
NPMTSearch Strategy=Greedy2017.0623.62
Sequence-to-sequence with attentionSearch Strategy=Beam Search2017.0622.59
REINFORCE-criticDecoding method=beam search2016.0722.58
seq2seqKte=12016.0622.53
LL (Wiseman & Rush, 2016)Decoding strategy=Greedy2017.0622.53
Actor-CriticDecoding method=beam search2016.0722.45
DADKte=102016.0622.4
DADKte=52016.0622.25
REINFORCE-criticDecoding method=greedy search2016.0722.24
MIXERDecoding Strategy=Beam2018.0621.83
MIXERKte=102016.0621.83
MIXER (Ranzato et al., 2015)Decoding strategy=Beam Search2017.0621.83
MIXERKte=52016.0621.81
Actor-CriticDecoding method=greedy search2016.0721.66
LLDecoding method=beam search2016.0721.46
REINFORCEDecoding method=beam search2016.0721.35
Sequence-to-sequence with attentionSearch Strategy=Greedy2017.0621.26
REINFORCEDecoding method=greedy search2016.0720.92
MIXERDecoding Strategy=Greedy2018.0620.73
MIXERKte=12016.0620.73
MIXERDecoding method=greedy search, Reference=Ranzato et al., 20152016.0720.73