NICT_LOGO.JPG KYOTO-U_LOGO.JPG

WAT

The Workshop on Asian Translation
Evaluation Results

[EVALUATION RESULTS TOP] | [BLEU] | [RIBES] | [AMFM] | [HUMAN (WAT2022)] | [HUMAN (WAT2021)] | [HUMAN (WAT2020)] | [HUMAN (WAT2019)] | [HUMAN (WAT2018)] | [HUMAN (WAT2017)] | [HUMAN (WAT2016)] | [HUMAN (WAT2015)] | [HUMAN (WAT2014)] | [EVALUATION RESULTS USAGE POLICY]

BLEU


# Team Task Date/Time DataID BLEU
Method
Other
Resources
System
Description
juman kytea mecab moses-
tokenizer
stanford-
segmenter-
ctb
stanford-
segmenter-
pku
indic-
tokenizer
unuse myseg kmseg
1cvitHINDENen-hi2020/07/06 19:08:493427------19.83---NMTYesMultilingual Transformer model.
2cvitHINDENen-hi2020/07/06 19:22:293428------20.52---NMTYesMultilingual Transformer model. Uses pib-v0 data.
3cvitHINDENen-hi2020/07/10 04:40:193436------20.69---NMTYesMultilingual model, uses pib-v2 data
4WTHINDENen-hi2020/09/03 18:20:143639------22.08---NMTNoUsed 5M Back translation news crawl data to train. Method: Transformer NMT; Preprocessing: 1. Removed mixed language sentences 2. moses tokeniser for English and for Hindi indicnlp normaliser and toke
5WTHINDENen-hi2020/09/03 23:14:363640------22.80---NMTYesMultilingual-ensembleX3
6NICT-5HINDENen-hi2020/09/18 17:47:183935------20.48---NMTNoMBART Fine Tune on approx. 900k sentence pairs from whole HindEn dataset.
7NICT-5HINDENen-hi2021/03/17 22:51:474557------19.00---NMTNoEnHi nmt model trained using my own toolkit. Only the parallel corpus is used. No fine tuning no pretraining. beam 4 lp 1.0.
8NICT-5HINDENen-hi2021/03/18 23:06:494571------19.42---NMTYesFT on an mBART model. Beam size 8.
9ORGANIZERHINDENen-hi2016/07/26 10:07:481032------18.72 0.00 0.00 0.00OtherYesOnline A (2016)
10ORGANIZERHINDENen-hi2016/07/26 13:24:221047------16.97 0.00 0.00 0.00OtherYesOnline B (2016)
11EHRHINDENen-hi2016/08/17 14:30:081166------11.75 0.00 0.00 0.00SMTNoPBSMT with preordering (DL=6)
12IITP-MTHINDENen-hi2016/08/18 23:13:251185------13.71 0.00 0.00 0.00SMTYesIITP-MT System1
13ORGANIZERHINDENen-hi2016/08/20 17:41:361252------10.79 0.00 0.00 0.00SMTNoPhrase-based SMT
14IITP-MTHINDENen-hi2016/08/29 18:51:441290------13.57 0.00 0.00 0.00SMTNoIITP-MT System2
15XMUNLPHINDENen-hi2017/07/20 23:07:381422------13.69 0.00 0.00 0.00NMTNosingle nmt model
16XMUNLPHINDENen-hi2017/07/27 22:04:541508------19.79 0.00 0.00 0.00NMTNosingle nmt model + monolingual data
17XMUNLPHINDENen-hi2017/07/28 23:38:291576------21.39 0.00 0.00 0.00NMTNoensemble of 4 nmt models + monolingual data
18IITB-MTGHINDENen-hi2017/08/01 15:09:011725------12.23 0.00 0.00 0.00NMTNoNMT with ensemble (last 3 + best validation)
19IITB-MTGHINDENen-hi2017/09/05 23:04:581763------ 0.34 0.00 0.00 0.00NMTNo
20cvitHINDENen-hi2018/09/07 12:29:042235------18.77- 0.00 0.00NMTYesConvS2S Model. External Data is used.
21cvitHINDENen-hi2018/09/09 01:20:092251------16.77- 0.00 0.00NMTNoConvS2S Model. IIT-Bombay data filtered with langdetect. + Backtranslated Monolingual Data ppl in [0.05, 0.14]
22cvitHINDENen-hi2018/09/09 21:12:292254------19.69- 0.00 0.00NMTYesConvS2S. Uses external data.
23CUNIHINDENen-hi2018/09/13 22:18:142320------19.78- 0.00 0.00NMTNoBig Transformer model with backtranslation, with transfer learning from English to Czech.
24CUNIHINDENen-hi2018/09/15 01:12:402361------20.28- 0.00 0.00NMTNoTransformer big, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 1300k steps
25CUNIHINDENen-hi2018/09/15 01:14:342362------17.63- 0.00 0.00NMTNoTransformer big, transfer learning from EN-CS 1M steps, followed by only backtranslation EN-HI for 300k steps, followed by original EN-HI for 500k steps, beam=8; alpha=0.8; averaging of last 8 models.
26CUNIHINDENen-hi2018/09/15 01:19:042363------16.49- 0.00 0.00NMTNoTransformer big, transfer learning from EN-CS 1M steps, only original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 230k steps.
27CUNIHINDENen-hi2018/09/15 01:20:332364------14.20- 0.00 0.00NMTNoBaseline, transformer big only EN-HI, beam=8, alpha=0.8, averaging 8 steps after 330k steps
28CUNIHINDENen-hi2018/09/15 01:22:032365------20.07- 0.00 0.00NMTNoTransformer big, transfer learning from EN-CS 1M steps, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 700k steps
29cvitHINDENen-hi2018/09/18 15:21:132489------21.10- 0.00 0.00NMTYesTransformer Base. Uses External Data
30cvitHINDENen-hi2018/09/18 21:58:212496------21.35- 0.00 0.00NMTYesTransformer Base. Uses External Data. Averaging of Checkpoints Enabled.
31cvitHINDENen-hi2018/09/18 22:37:372500------21.57- 0.00 0.00NMTYesAveraging Models from epochs 61-68. Base Transformer. Uses External Data.
32ORGANIZERHINDENen-hi2018/11/13 14:54:582566------13.76- 0.00 0.00NMTNoNMT with Attention
33cvitHINDENen-hi2019/03/15 01:21:272642------20.17- 0.00 0.00NMTYesmassive-multi
34cvitHINDENen-hi2019/03/15 01:31:412644------18.31- 0.00 0.00NMTYesmassive-multi + ft
35cvitHINDENen-hi2019/05/27 16:03:362680------20.46- 0.00 0.00NMTYesmassive-multi + bt

Notice:

Back to top

RIBES


# Team Task Date/Time DataID RIBES
Method
Other
Resources
System
Description
juman kytea mecab moses-
tokenizer
stanford-
segmenter-
ctb
stanford-
segmenter-
pku
indic-
tokenizer
unuse myseg kmseg
1cvitHINDENen-hi2020/07/06 19:08:493427------0.758405---NMTYesMultilingual Transformer model.
2cvitHINDENen-hi2020/07/06 19:22:293428------0.766753---NMTYesMultilingual Transformer model. Uses pib-v0 data.
3cvitHINDENen-hi2020/07/10 04:40:193436------0.764496---NMTYesMultilingual model, uses pib-v2 data
4WTHINDENen-hi2020/09/03 18:20:143639------0.765340---NMTNoUsed 5M Back translation news crawl data to train. Method: Transformer NMT; Preprocessing: 1. Removed mixed language sentences 2. moses tokeniser for English and for Hindi indicnlp normaliser and toke
5WTHINDENen-hi2020/09/03 23:14:363640------0.769138---NMTYesMultilingual-ensembleX3
6NICT-5HINDENen-hi2020/09/18 17:47:183935------0.763000---NMTNoMBART Fine Tune on approx. 900k sentence pairs from whole HindEn dataset.
7NICT-5HINDENen-hi2021/03/17 22:51:474557------0.750840---NMTNoEnHi nmt model trained using my own toolkit. Only the parallel corpus is used. No fine tuning no pretraining. beam 4 lp 1.0.
8NICT-5HINDENen-hi2021/03/18 23:06:494571------0.757646---NMTYesFT on an mBART model. Beam size 8.
9ORGANIZERHINDENen-hi2016/07/26 10:07:481032------0.7167880.0000000.0000000.000000OtherYesOnline A (2016)
10ORGANIZERHINDENen-hi2016/07/26 13:24:221047------0.6912980.0000000.0000000.000000OtherYesOnline B (2016)
11EHRHINDENen-hi2016/08/17 14:30:081166------0.6718660.0000000.0000000.000000SMTNoPBSMT with preordering (DL=6)
12IITP-MTHINDENen-hi2016/08/18 23:13:251185------0.6889130.0000000.0000000.000000SMTYesIITP-MT System1
13ORGANIZERHINDENen-hi2016/08/20 17:41:361252------0.6511660.0000000.0000000.000000SMTNoPhrase-based SMT
14IITP-MTHINDENen-hi2016/08/29 18:51:441290------0.6830220.0000000.0000000.000000SMTNoIITP-MT System2
15XMUNLPHINDENen-hi2017/07/20 23:07:381422------0.7128760.0000000.0000000.000000NMTNosingle nmt model
16XMUNLPHINDENen-hi2017/07/27 22:04:541508------0.7431290.0000000.0000000.000000NMTNosingle nmt model + monolingual data
17XMUNLPHINDENen-hi2017/07/28 23:38:291576------0.7496600.0000000.0000000.000000NMTNoensemble of 4 nmt models + monolingual data
18IITB-MTGHINDENen-hi2017/08/01 15:09:011725------0.6886060.0000000.0000000.000000NMTNoNMT with ensemble (last 3 + best validation)
19IITB-MTGHINDENen-hi2017/09/05 23:04:581763------0.3012410.0000000.0000000.000000NMTNo
20cvitHINDENen-hi2018/09/07 12:29:042235------0.748008-0.0000000.000000NMTYesConvS2S Model. External Data is used.
21cvitHINDENen-hi2018/09/09 01:20:092251------0.714197-0.0000000.000000NMTNoConvS2S Model. IIT-Bombay data filtered with langdetect. + Backtranslated Monolingual Data ppl in [0.05, 0.14]
22cvitHINDENen-hi2018/09/09 21:12:292254------0.758365-0.0000000.000000NMTYesConvS2S. Uses external data.
23CUNIHINDENen-hi2018/09/13 22:18:142320------0.754244-0.0000000.000000NMTNoBig Transformer model with backtranslation, with transfer learning from English to Czech.
24CUNIHINDENen-hi2018/09/15 01:12:402361------0.761292-0.0000000.000000NMTNoTransformer big, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 1300k steps
25CUNIHINDENen-hi2018/09/15 01:14:342362------0.753895-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, followed by only backtranslation EN-HI for 300k steps, followed by original EN-HI for 500k steps, beam=8; alpha=0.8; averaging of last 8 models.
26CUNIHINDENen-hi2018/09/15 01:19:042363------0.754966-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, only original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 230k steps.
27CUNIHINDENen-hi2018/09/15 01:20:332364------0.733738-0.0000000.000000NMTNoBaseline, transformer big only EN-HI, beam=8, alpha=0.8, averaging 8 steps after 330k steps
28CUNIHINDENen-hi2018/09/15 01:22:032365------0.761582-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 700k steps
29cvitHINDENen-hi2018/09/18 15:21:132489------0.771549-0.0000000.000000NMTYesTransformer Base. Uses External Data
30cvitHINDENen-hi2018/09/18 21:58:212496------0.773078-0.0000000.000000NMTYesTransformer Base. Uses External Data. Averaging of Checkpoints Enabled.
31cvitHINDENen-hi2018/09/18 22:37:372500------0.773923-0.0000000.000000NMTYesAveraging Models from epochs 61-68. Base Transformer. Uses External Data.
32ORGANIZERHINDENen-hi2018/11/13 14:54:582566------0.710210-0.0000000.000000NMTNoNMT with Attention
33cvitHINDENen-hi2019/03/15 01:21:272642------0.761061-0.0000000.000000NMTYesmassive-multi
34cvitHINDENen-hi2019/03/15 01:31:412644------0.718374-0.0000000.000000NMTYesmassive-multi + ft
35cvitHINDENen-hi2019/05/27 16:03:362680------0.765422-0.0000000.000000NMTYesmassive-multi + bt

Notice:

Back to top

AMFM


# Team Task Date/Time DataID AMFM
Method
Other
Resources
System
Description
unuse unuse unuse unuse unuse unuse unuse unuse unuse unuse
1cvitHINDENen-hi2020/07/06 19:08:493427------0.867360---NMTYesMultilingual Transformer model.
2cvitHINDENen-hi2020/07/06 19:22:293428------0.866410---NMTYesMultilingual Transformer model. Uses pib-v0 data.
3cvitHINDENen-hi2020/07/10 04:40:193436------0.868770---NMTYesMultilingual model, uses pib-v2 data
4WTHINDENen-hi2020/09/03 18:20:143639------0.869400---NMTNoUsed 5M Back translation news crawl data to train. Method: Transformer NMT; Preprocessing: 1. Removed mixed language sentences 2. moses tokeniser for English and for Hindi indicnlp normaliser and toke
5WTHINDENen-hi2020/09/03 23:14:363640------0.873830---NMTYesMultilingual-ensembleX3
6NICT-5HINDENen-hi2020/09/18 17:47:183935------0.864600---NMTNoMBART Fine Tune on approx. 900k sentence pairs from whole HindEn dataset.
7NICT-5HINDENen-hi2021/03/17 22:51:474557------0.861390---NMTNoEnHi nmt model trained using my own toolkit. Only the parallel corpus is used. No fine tuning no pretraining. beam 4 lp 1.0.
8NICT-5HINDENen-hi2021/03/18 23:06:494571------0.861970---NMTYesFT on an mBART model. Beam size 8.
9ORGANIZERHINDENen-hi2016/07/26 10:07:481032------0.6706600.0000000.0000000.000000OtherYesOnline A (2016)
10ORGANIZERHINDENen-hi2016/07/26 13:24:221047------0.6684500.0000000.0000000.000000OtherYesOnline B (2016)
11EHRHINDENen-hi2016/08/17 14:30:081166------0.6507500.0000000.0000000.000000SMTNoPBSMT with preordering (DL=6)
12IITP-MTHINDENen-hi2016/08/18 23:13:251185------0.6573300.0000000.0000000.000000SMTYesIITP-MT System1
13ORGANIZERHINDENen-hi2016/08/20 17:41:361252------0.6608600.0000000.0000000.000000SMTNoPhrase-based SMT
14IITP-MTHINDENen-hi2016/08/29 18:51:441290------0.6632100.0000000.0000000.000000SMTNoIITP-MT System2
15XMUNLPHINDENen-hi2017/07/20 23:07:381422------0.6477400.0000000.0000000.000000NMTNosingle nmt model
16XMUNLPHINDENen-hi2017/07/27 22:04:541508------0.6825000.0000000.0000000.000000NMTNosingle nmt model + monolingual data
17XMUNLPHINDENen-hi2017/07/28 23:38:291576------0.6887700.0000000.0000000.000000NMTNoensemble of 4 nmt models + monolingual data
18IITB-MTGHINDENen-hi2017/08/01 15:09:011725------0.6247800.0000000.0000000.000000NMTNoNMT with ensemble (last 3 + best validation)
19IITB-MTGHINDENen-hi2017/09/05 23:04:581763------0.4633500.0000000.0000000.000000NMTNo
20cvitHINDENen-hi2018/09/07 12:29:042235------0.697630-0.0000000.000000NMTYesConvS2S Model. External Data is used.
21cvitHINDENen-hi2018/09/09 01:20:092251------0.664330-0.0000000.000000NMTNoConvS2S Model. IIT-Bombay data filtered with langdetect. + Backtranslated Monolingual Data ppl in [0.05, 0.14]
22cvitHINDENen-hi2018/09/09 21:12:292254------0.699810-0.0000000.000000NMTYesConvS2S. Uses external data.
23CUNIHINDENen-hi2018/09/13 22:18:142320------0.700240-0.0000000.000000NMTNoBig Transformer model with backtranslation, with transfer learning from English to Czech.
24CUNIHINDENen-hi2018/09/15 01:12:402361------0.704220-0.0000000.000000NMTNoTransformer big, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 1300k steps
25CUNIHINDENen-hi2018/09/15 01:14:342362------0.693830-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, followed by only backtranslation EN-HI for 300k steps, followed by original EN-HI for 500k steps, beam=8; alpha=0.8; averaging of last 8 models.
26CUNIHINDENen-hi2018/09/15 01:19:042363------0.690150-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, only original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 230k steps.
27CUNIHINDENen-hi2018/09/15 01:20:332364------0.681460-0.0000000.000000NMTNoBaseline, transformer big only EN-HI, beam=8, alpha=0.8, averaging 8 steps after 330k steps
28CUNIHINDENen-hi2018/09/15 01:22:032365------0.701300-0.0000000.000000NMTNoTransformer big, transfer learning from EN-CS 1M steps, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 700k steps
29cvitHINDENen-hi2018/09/18 15:21:132489------0.712200-0.0000000.000000NMTYesTransformer Base. Uses External Data
30cvitHINDENen-hi2018/09/18 21:58:212496------0.712010-0.0000000.000000NMTYesTransformer Base. Uses External Data. Averaging of Checkpoints Enabled.
31cvitHINDENen-hi2018/09/18 22:37:372500------0.712110-0.0000000.000000NMTYesAveraging Models from epochs 61-68. Base Transformer. Uses External Data.
32ORGANIZERHINDENen-hi2018/11/13 14:54:582566------0.644860-0.0000000.000000NMTNoNMT with Attention
33cvitHINDENen-hi2019/03/15 01:21:272642------0.701670-0.0000000.000000NMTYesmassive-multi
34cvitHINDENen-hi2019/03/15 01:31:412644------0.680620-0.0000000.000000NMTYesmassive-multi + ft
35cvitHINDENen-hi2019/05/27 16:03:362680------0.702380-0.0000000.000000NMTYesmassive-multi + bt

Notice:

Back to top

HUMAN (WAT2022)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description

Notice:
Back to top

HUMAN (WAT2021)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description

Notice:
Back to top

HUMAN (WAT2020)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description
1WTHINDENen-hi2020/09/03 23:14:3636403.560NMTYesMultilingual-ensembleX3
2WTHINDENen-hi2020/09/03 18:20:1436393.490NMTNoUsed 5M Back translation news crawl data to train. Method: Transformer NMT; Preprocessing: 1. Removed mixed language sentences 2. moses tokeniser for English and for Hindi indicnlp normaliser and toke

Notice:
Back to top

HUMAN (WAT2019)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description
1cvitHINDENen-hi2019/05/27 16:03:362680UnderwayNMTYesmassive-multi + bt

Notice:
Back to top

HUMAN (WAT2018)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description
1CUNIHINDENen-hi2018/09/15 01:14:34236277.000NMTNoTransformer big, transfer learning from EN-CS 1M steps, followed by only backtranslation EN-HI for 300k steps, followed by original EN-HI for 500k steps, beam=8; alpha=0.8; averaging of last 8 models.
2cvitHINDENen-hi2018/09/09 21:12:29225469.500NMTYesConvS2S. Uses external data.
3CUNIHINDENen-hi2018/09/15 01:22:03236560.000NMTNoTransformer big, transfer learning from EN-CS 1M steps, only backtranslation EN-HI, no original EN-HI, beam=8; alpha=0.8; averaging of last 8 models after 700k steps
4cvitHINDENen-hi2018/09/09 01:20:09225150.500NMTNoConvS2S Model. IIT-Bombay data filtered with langdetect. + Backtranslated Monolingual Data ppl in [0.05, 0.14]
5cvitHINDENen-hi2018/09/07 12:29:042235UnderwayNMTYesConvS2S Model. External Data is used.

Notice:
Back to top

HUMAN (WAT2017)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description
1XMUNLPHINDENen-hi2017/07/28 23:38:29157664.500NMTNoensemble of 4 nmt models + monolingual data
2IITB-MTGHINDENen-hi2017/08/01 15:09:01172528.750NMTNoNMT with ensemble (last 3 + best validation)

Notice:
Back to top

HUMAN (WAT2016)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description
1ORGANIZERHINDENen-hi2016/07/26 10:07:48103257.250OtherYesOnline A (2016)
2ORGANIZERHINDENen-hi2016/07/26 13:24:22104742.500OtherYesOnline B (2016)
3IITP-MTHINDENen-hi2016/08/18 23:13:2511854.750SMTYesIITP-MT System1
4EHRHINDENen-hi2016/08/17 14:30:0811660.000SMTNoPBSMT with preordering (DL=6)

Notice:
Back to top

HUMAN (WAT2015)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description

Notice:
Back to top

HUMAN (WAT2014)


# Team Task Date/Time DataID HUMAN
Method
Other
Resources
System
Description

Notice:
Back to top

EVALUATION RESULTS USAGE POLICY

When you use the WAT evaluation results for any purpose such as:
- writing technical papers,
- making presentations about your system,
- advertising your MT system to the customers,
you can use the information about translation directions, scores (including both automatic and human evaluations) and ranks of your system among others. You can also use the scores of the other systems, but you MUST anonymize the other system's names. In addition, you can show the links (URLs) to the WAT evaluation result pages.

NICT (National Institute of Information and Communications Technology)
Kyoto University
Last Modified: 2018-08-02