Interested in Contributing?

Scored Systems

System Submitter System Notes Constraint Run Notes BLEU BLEU-cased TER BEER 2.0 CharactTER
tilde-c-nmt-1bt  (Details) m4t1ss
Tilde
6-layer encoder, 6-layer decoder transformer model trained on filtered parallel and filtered back-translated mono data yes Post WMT17

22.2

21.7

0.674

0.572

0.540

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes

21.1

20.7

0.676

0.565

0.554

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes HNMT with hybrid encoder and BPE decoder, backtranslated training data, ensemble of 4 systems with averaging of 4 savepoints each, with dash re-ranking

21.0

20.6

0.680

0.563

0.557

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes

20.6

20.3

0.687

0.561

0.560

AaltoHnmtMultitask  (Details) Stig-Arne Grönroos
Aalto University
yes HNMT with hybrid word/character decoder, multi-task learning of FinnPos tags and clustered lemmas, beam search with penalties, back-translated monolingual data, ensemble of 4.

20.6

20.3

0.673

0.554

0.592

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes HNMT with hybrid encoder and BPE decoder, backtranslated training data, ensemble of 4 systems with averaging of 4 savepoints each, without dash re-ranking

20.1

19.7

0.712

0.560

0.565

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes ... after the deadline (combination with PBSMT)

19.8

19.3

0.698

0.559

0.564

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes character-based model

19.5

19.1

0.685

0.559

0.565

AaltoHnmtFlatcat  (Details) Stig-Arne Grönroos
Aalto University
yes HNMT with Morfessor FlatCat subwords, back-translated monolingual data, ensemble of 4.

17.4

17.2

0.750

0.528

0.600

HY-SMT-OPUS  (Details) jorgtied
University of Helsinki
phrase-based SMT with additional data from OPUS no all WMT data, back-translated news, OPUS, OSM

17.3

17.0

0.735

0.553

0.589

HY-HNMT  (Details) jorgtied
University of Helsinki
Helsinki Neural Machine Translation system (HNMT) yes training data combined with SMT translations, BPE on both sides

17.2

16.8

0.746

0.544

0.588

jhu_nmt_lattice_rescore  (Details) huda
Johns Hopkins Univesity
yes NMT rescoring of PBMT lattice; pruning threshold of .5; stack size of 100;

16.4

16.0

0.758

failed

failed

HY-SMT  (Details) jorgtied
University of Helsinki
standard phrase-based Moses system yes phrase-based SMT with BPE, back-translated news, CC-LM

16.2

15.9

0.754

0.537

0.621

Moses phrase-based, word clusters  (Details) jhu-smt
Johns Hopkins University
yes Moses Phrase-Based, word cluster language models

14.9

14.5

0.778

0.523

0.628

jhu_nmt_nbest_rescore  (Details) gkumar
Johns Hopkins University
PBMT 500 best rescored with NMT-ensemble. yes PBMT 500 best rescored with NMT-ensemble.

15.4

14.3

0.774

0.520

0.616

ParFDA  (Details) bicici
en-fi ParFDA Moses phrase-based SMT system yes en-fi (after the deadline)

13.2

12.8

0.793

0.508

0.658

TALP-UPC  (Details) cescolano
TALP-UPC
Character to character NMT system with rescoring using the inverse language pair model. yes

11.8

11.5

0.806

0.483

0.714

HY-AH  (Details) jorgtied
University of Helsinki
handcrafted rule-based system by Arvi Hurskainen no

9.8

9.3

0.857

0.493

0.620

apertium-fin-eng-unconstrained-en-fi-overfit  (Details) tpirinen
This is a rule-based MT and test set included in dictionary. no Added lexical selection rules statistically.

3.5

3.4

1.027

0.403

0.759

apertium-fin-eng-unconstrained-en-fi-overfit  (Details) tpirinen
This is a rule-based MT and test set included in dictionary. no fixed off-by-one line wrap problem!

3.5

3.4

1.028

0.400

0.764

apertium-fin-eng-unconstrained-en-fi  (Details) tpirinen
This is RBMT with large monolingual and bilingual dictionaries but minimal amount of rules. This submission was made by "tuning" on dev sets, i.e. manually fixing system to be best on dev sets. no Some 70 % of dev set "tuned"

1.1

1.1

1.246

0.294

failed

vagrant lexicon  (Details) anastasia
no

failed

failed

failed

0.000

0.000