Machine Translation On Wmt2014 English German

평가 지표

BLEU score

Hardware Burden

Operations per network pass

평가 결과

이 벤치마크에서 각 모델의 성능 결과

모델 이름	BLEU score	Hardware Burden	Operations per network pass	Paper Title	Repository
Transformer Big + adversarial MLE	29.52			Improving Neural Language Modeling via Adversarial Training
MAT	-	-	-	Multi-branch Attentive Transformer
AdvAug (aut+adv)	29.57	-	-	AdvAug: Robust Adversarial Augmentation for Neural Machine Translation	-
CMLM+LAT+4 iterations	27.35			Incorporating a Local Translation Mechanism into Non-autoregressive Translation	-
FlowSeq-large (IWD n = 15)	22.94			FlowSeq: Non-Autoregressive Conditional Sequence Generation with Generative Flow
Transformer (ADMIN init)	30.1	-	-	Very Deep Transformers for Neural Machine Translation
MUSE(Parallel Multi-scale Attention)	29.9			MUSE: Parallel Multi-Scale Attention for Sequence to Sequence Learning
Transformer-DRILL Base	28.1			Deep Residual Output Layers for Neural Language Generation
Transformer Big with FRAGE	29.11			FRAGE: Frequency-Agnostic Word Representation
GLAT	25.21	-	-	Glancing Transformer for Non-Autoregressive Neural Machine Translation
PartialFormer	29.56	-	-	PartialFormer: Modeling Part Instead of Whole for Machine Translation
Bi-SimCut	30.78	-	-	Bi-SimCut: A Simple Strategy for Boosting Neural Machine Translation
Transformer + SRU	28.4	34G		Simple Recurrent Units for Highly Parallelizable Recurrence
PBMT	20.7			-	-
Local Joint Self-attention	29.7			Joint Source-Target Self Attention with Locality Constraints
Lite Transformer	26.5	-	-	Lite Transformer with Long-Short Range Attention
Average Attention Network (w/o FFN)	26.05	-	-	Accelerating Neural Transformer via an Average Attention Network
Unsupervised NMT + Transformer	17.16			Phrase-Based & Neural Unsupervised Machine Translation
KERMIT	28.7			KERMIT: Generative Insertion-Based Modeling for Sequences	-
T2R + Pretrain	28.7			Finetuning Pretrained Transformers into RNNs

0 of 91 row(s) selected.