1) The document discusses Pangeanic's new neural machine translation (NMT) technology called PangeaMT Neural and ActivaTM and compares its performance to their previous statistical machine translation (SMT) system.
2) Experimental results on English to Japanese translation show that NMT outperforms SMT in BLEU, TER, and WER scores, especially for shorter sentences between 0-25 words, producing translations requiring less post-editing effort.
3) Additional testing of NMT on other language pairs like English to French and Russian also showed superior results compared to SMT, with translators rating 85-90% of NMT translations as good or very good quality versus only 50-60
Human Factors of XR: Using Human Factors to Design XR Systems
The Web, Database and Neural NMT Comparison
1. The Web, The Database and
The Neural
Garth Hedenskog, Sales Director
Pangeanic TAUS Girona, 13 June 2017
2. • National research project CDTI
• Workflow system with built in crawler
• PM-less track and workflows initiation
• Powerful tool with incorporated with
Pangeanic’s new technology – ActivaTM and
PangeaMT Neural
3. ELASTIC CENTRALIZED TM SYSTEM
• FEATURES:
• CAT tool agnostic
• Cor integratable
• Hosting options
• Tag handling capabilities
• API to NMT
• Triangulation
…summary
4. Our story……
• First translation company in the world to make commercial use of
Moses.
• Wins a post-editing contract in 2007 to work for the European
Commission as MT output post-editors.
THAT WAS THEN, THIS IS NOW
• Pangeanic’s consortium, along with KantanMT, Prompsit and Tilde,
was awarded the largest EU contract by CEF (Connecting Europe
Facility) to supply infrastructure services to the European Union in
the field of Digital Service Infrastructures, and particularly machine
translation. (IADAATPA (Intelligent, Automatic Domain Adapted
Automated Translation for Public Administrations)
5. Training Corpus
Sentences Running
words
Vocabulary
EN 4,6M 55,9M 491,6K
JA 4,6M 76,0M 283,8K
Dev corpus
Sentences Running
words
OOVs
EN 1,9K 24,1K 1,32
JA 1,9K 32,7K 0,86
Test corpus
Sentences Running
words
OOVs Average length
in characters
Average
number tokens
EN 2K 27,1K 1,80 77 14,12
JA 2K 37,0K 1,14 59 19,08
Training data:
• TAUS data for Electronics Computer Hardware (ECH) plus SOFT (IT) 4,6M sentences / 56M words (EN)
• EN and JA tokenized (tokenizer.perl and Mecab respectively)
BLEU TER WER
PangeaMT 43,25 0,493174 0,607223
NMT 44,53 0,422858 0,473214
Seemingly…. Not such a big difference
Results EN->JP :
6. 0-10 words 11-15 words 16-20 words 21-25 words 26-30 words 31+ words
BLEU TER WER BLEU TER WER BLEU TER WER BLEU TER WER BLEU TER WER BLEU TER WER
Pangea
MT
44,00 0,428
65
0,471
268
42,80 0,465
28
0,591
708
41,08 0,485
096
0,617
126
39,95 0,491
183
0,649
891
39,08 0,539
768
0,693
745
35,38 0,565
217
0,713
226
NMT 40,59 0,398
68
0,414
078
46,00 0,353
941
0,393
642
43,43 0,392
998
0,443
898
42,04 0,407
965
0,476
323
39,86 0,461
081
0,529
578
35,65 0,561
833
0,630
695
Results EN->JP by length:
• In shorter sentences (0-10 words), our SMT system scores better results in BLEU, but if we take a look to the
TER and WER, we see that in character and word level, NMT has better results which means less post edition
efforts.In sentences (11-25 words), NMT always gets better results in BLEU, WER and TER.
• In longer sentences (26++), NMT tends to have same results than PangeaMT.
BLEU TER WER
7. Tests in F/I/G/S, RU, PT point to a very strong preference towards NMT (results available in our blog)
On average: from a set of (random) 250 sentences, around 85% - 90%, were good or very good (A or B). ES/PT/IT
results similar to FR
Evaluation: Translation companies and professional freelance translators
EN-DE set of 250 sentences
NMT SMT
A 132 53% 34 14%
B 98 39% 95 38%
C 14 6% 97 39%
D 6 2% 24 10%
250 250
EN-FR set of 250 sentences
NMT SMT
A 150 60% 39 16%
B 76 30% 126 50%
C 21 8% 71 28%
D 3 1% 14 6%
250 250
EN-RU set of 250 sentences
NMT SMT
A 128 51% 39 16%
B 84 34% 43 17%
C 22 9% 60 24%
D 16 6% 108 43%
250 250
EN-JP set of 250 sentences
NMT SMT
A 83 33% 17 7%
B 71 28% 14 6%
C 56 22% 95 38%
D 40 16% 124 50%
250 250
8. •Conclusion
•NN does not produce miracles yet but the innitial results are very exciting.
•The shift is remarkable in all languages especially JP which has moved away from
the usual average to bad results to a great leap to pretty acceptable quality
Thank you!
garth@pangeanic.com