2013 ALC Conference

10 Decisions to make before starting to use Machine Translation (MT).

Details on how to improve MT engines.

Transcript

1. Your Trained Moses SMT System doesn't work. What can you do? Diego Bartolome, CEO tauyou <language technology> diego.bartolome@tauyou.com @diegobartolome
2. Where are you now?
3. Where are you now?
4. Why Machine Translation? Strategic decision Increase sales Shorten delivery times Reduce costs Differentiation Forced decision Clients ask for it!
5. Dare – change 5
6. Welcome to the jungle
7. Decision 1: Internal – external Core competence Resources ROI Time to market
8. Decision 1: Internal – external Core competence Resources ROI Time to market
9. MT Costs Internal development Free tools DOiY solutions Traditional pricing model tauyou managed solution
10. Decision 2: MT Type (I) Rule-based MT Statistical MT Hybrid MT
11. Decision 2: MT Type (II) Do we really care?
12. Decision 3: Languages (I) Source: translate.autodesk.com
13. Decision 3: Languages (II) Source: Philipp Koehn
14. Decision 4: Domains Who is willing to pay? Where does your revenue come from? What are your key skills? What domains achieve good quality?
15. Decision 5: Workflow Use MT as a secondary TM Bilingual pre-translated translation files CAT tool integration Differentiated workflow
16. Decision 6: Feedback Qualitative Use updated TMs in new trainings Immediate (incremental) retraining Rule-based automatic post-editing Selective pre- and/or post-processing
17. Decision 7: Post-editors What are the skills needed? Post-editing guidelines How do we pay them?
18. Decision 8: Metrics SMT metrics: BLEU, NIST Feedback from translators Translation time vs. Post-editing time Word Error Rate (WER) or Edit Distance Cost reduction
19. Decision 9: Business Model
21. Decision 10: Start!
22. Let's play with Moses
23. Let's play with Moses Best resource to start www.statmt.org/moses TAUS tutorial www.translationautomation.com tauyou slides www.speakerdeck.com/tauyoucom
24. Everything is clear! Gather TMs and other linguistic assets Select domains Train systems BLEU score is great … but … Translation quality is awful
25. Why? Not enough data Too much data Unclean TMs Misalignments Difficult language pairs Selection of wrong parameters Suboptimal techniques
27. Some steps Maximum exploitation of existing assets Source content optimization Data selection and cleaning Improvement of the models Linguistic processing Continuous improvement
28. Linguistic assets Translation memory sharing Clients, Partners, EU, UN, TAUS Relevant on-line data retrieval Advanced TM techniques Sub-segment matching Parts of Speech replacement
29. Source optimization (I) Spell check Grammar check Style check Terminology check Client checklist new doc proposed doc + html report
30. Summarization % to reduce Use translation memories Project Client All new doc proposed doc + html report
31. Data selection + cleaning Clean translation memories Length, punctuation, terminology, … Inconsistencies, repetitions, ... Segment splitting Optimize weight of most frequent n-grams Validate their translations Add out-of-domain data
32. Models optimization Filter the translation tables Remove the garbage + tune weights Optimize language models Adapt them to the translation purpose Tune parameters correctly Tune set, test set, optimization parameters Improve recasing
33. Linguistic processing In the source and/or target language Grammar checking Entities detection Proper nouns, alphanumeric words, ... Compound words splitting Sentence reordering
34. Life is about the people you meet and the things you create with them. So go out and start creating Part of the Holstee Manifesto Diego Bartolome CEO tauyou <language technology> diego.bartolome@tauyou.com @diegobartolome
35. Thank you! Diego Bartolome CEO tauyou <language technology> diego.bartolome@tauyou.com @diegobartolome