Lexcelera MT Breaking Compromises
- 1. Breaking Compromises Optimized MT for Price, Speed and Quality
- 2. Different approaches to MT Factors driving MT today Changing the current localization paradigm MT breaks the price-quality compromises Optimizing MT quality Lexcelera’s MT services Keys to MT success
- 3. I. Different Approaches to MT Machine Translation is not one program …but different programs taking different approaches to get to the same place . Different approaches to MT
- 4. Rules-Based MT (RMBT) e.g. Systran, ProMT Comes with linguistic rules, trained by domain-specific user dictionaries . Different approaches to MT
- 5. Statistical MT (SMT) e.g. Language Weaver, Google, Asia Online, Moses (open source) May have domain built in, trained by millions of segments . Different approaches to MT
- 6. Advantages of RBMT ▪ Customization requires trained linguists ▪ Only one language pair at a time ▪ Not available for every language ▪ Labour intensive to scale/extend to new domains (4-6 weeks) ▪ Respects grammatical rules: Le chat vert ▪ User dictionaries control terms ▪ System can be trained/corrected in real time ▪ Large corpora not necessary ▪ Most user cases today are Systran Disadvantages of RBMT . Different approaches to MT
- 7. Advantages of SMT ▪ Requires millions of segments of clean training text (TMs, bilingual and monolingual corpora) ▪ Unpredictable: Terms may be correct in one sentence and not in another ▪ System makes its own decisions based on probability (e.g. 75% likelihood that “banco” = “bank” and not “bench” so always chooses more likely) ▪ Corrections not always easy to integrate into engine ▪ Memory and processor intensive: based on server licences ▪ No grammatical rules to govern text it hasn’t seen: le vert chat ▪ Learns automatically ▪ Same process for virtually any language ▪ Customization can be done with limited human input ▪ Extending to new languages/domains takes up processing capacity, not linguistic resources (weeks to months of training) Disadvantages of SMT . Different approaches to MT
- 8. Systran introduced Hybrid RBMT/SMT in May 2009 SMT systems seek to add RBMT module A Hybrid RBMT/SMT System? . Different approaches to MT
- 9. More content to translate More languages Need for speed to reach global markets Budget cuts Process retooling II. Factors Driving MT Today
- 10. ▪ Alerts ▪ Customer support ▪ Intelligence/analytics ▪ Help/documentation ▪ Continuous updates and fixes ▪ Courseware ▪ Company intranets ▪ Knowledge bases ▪ User-generated content More content to translate . . Factors driving MT today
- 11. ▪ C ustomers more likely to buy in their own language ▪ Customers prefer support in their own language ▪ Global employees follow instructions better in their own language ▪ Merged enterprises need terminology and infrastructure standardized across global entities More languages . . Factors driving MT today
- 12. Need for speed: Continuous updates, simship, online customer support, alerts, time to market… . Factors driving MT today
- 13. Today companies are re-engineering processes to increase speed and capacity, and reduce costs. . Factors driving MT today “ Translate more content, more languages, do it faster… and for less.” Budget cuts .
- 14. The strongest companies will emerge from the global recession able to manage more languages and more content at faster speeds and for less cost . Factors driving MT today To do this, they need a new paradigm to break the traditional price vs quality compromise
- 15. III. Changing the Localization Paradigm If you want speed and quality you have to compromise on price If you want speed and price you h a v e t o compromise on quality . Changing the localization paradigm The Current Paradigm Price Speed Quality Pick 2 only
- 16. It’s time to retire the current localization paradigm. . Changing the localization paradigm
- 17. IV. MT breaks the compromises Optimized MT can optimize speed, price and quality
- 18. ▪ Post-editors don’t need to search terms (if well trained) ▪ After training, up to 7000 words per day for human quality, 10,000 for “comprehensibility” ▪ Updates in real time ▪ Twice as fast to post-edit MT output than translate from scratch ▪ Siemens study: 6.28 seconds to translate no match segment 2.44 seconds to translate fuzzy match 3.40 seconds to post-edit MT segment . MT breaks the speed compromise Optimized MT can improve speed .
- 19. ▪ Bentley Systems Inc, an engineering software publisher serving more than 80% of the top 500 engineering design companies worldwide ▪ Software help/documentation project ▪ 1 million words in sgml files, 50% leveraging, ▪ 20 working days with 4.5 post-editors and 1 linguistic team leader ▪ Turnaround time decreased by one-third, including engineering Lexcelera Case Study . . MT breaks the speed compromise
- 20. ▪ Intel was able to use MT to push updates to its knowledge base out to customers 10 times more quickly, now publishing new articles in 24 hours, instead of 10 days. Intel Case Study . . MT breaks the speed compromise
- 21. ▪ Higher productivity post-editing MT than translating so word costs are significantly lower ▪ Terminology may be managed automatically (RBMT) ▪ Online knowledge bases, FAQs, translated in customers’ languages reduce customer support calls, emails ▪ ROI from higher customer satisfaction and increased local sales ▪ Lower internal headcount can manage higher translation volumes . MT breaks the cost compromise Optimized MT can lower costs considerably .
- 22. ▪ Veolia (CAC 40 company) required rapid translation, ‘understandable quality’ for the adaptation and deployment of SAP within Veolia. ▪ 44 Word documents 17 PowerPoint Presentations A total of more than 490,000 words ▪ 3 post-editors, 1 month Lexcelera Case Study . . MT breaks the cost compromise
- 23. ▪ Symantec built an in-house system to combine process automation with MT using Systran ▪ Productivity doubled ▪ Costs were cut by 30% ▪ Over three years, translation volumes (80% documentation) grew from 10 million words per year to 30 million without an increase in internal headcount ▪ The actual price of translation was halved in 2.5 years Symantec Case Study . . MT breaks the cost compromise
- 24. ▪ Pan American Health Organization (PAHO) reduced costs by 33% ▪ European Patent Office makes patent content available ▪ Microsoft, savings of 5% to 25% ▪ Cisco, savings of 50% on Japanese for knowledge base ▪ MBDA, 50-80% gain (useful quality, no post-editing) ▪ Siemens doubled productivity ▪ Intel achieved a call centre deflection rate of 10% Other Case Studies . . MT breaks the cost compromise
- 25. ▪ Once terminology is defined, the MT engine will apply consistent terminology across all documents ▪ MT erases stylistic/terminology differences among multiple translators ▪ MT optimizes TM leverage from previous translations ▪ Where localization is generally T+P, post-editors make the process T+E+P ▪ Virtuous circle of feedback improves the engine in real-time ▪ Quality improves as the system improves with each successive round of feedback is entered into the engine . MT breaks the quality compromise Optimized MT can improve quality .
- 26. ▪ “ Contrary to all expectations, using MT in Bentley has improved the translation quality in the pilot projects” ▪ Recent evaluation of German courseware translation: “It was the best translation of courseware I ever read.” Lexcelera Case Study . . MT breaks the quality compromise
- 27. ▪ Customers report greater satisfaction with the quality of machine translated texts than with no translation (e.g. information available only in the original language). ▪ Symantec: "In a customer care scenario, French and German customers found machine translated documents useful in 50% to 75% of the time.” ▪ Microsoft: on average, 23% found machine-translated articles to be useful vs. 29% for human translation. ▪ Intel: raw MT content was only 3% less successful in answering customer questions (44% vs. 47%) Other Case Studies: MT with no human input . . MT breaks the quality compromise
- 28. ▪ Microsoft measured higher rates of satisfaction for MT translations of their knowledge base without human input, as below, and concludes: “ It appears that the Spanish users were so happy to have all the articles in their own language that they were willing to overlook the fact that their quality was less than that of human translations. ” Microsoft Case Study . . MT breaks the quality compromise 87% 69% N/A 72% % of customers who thought information is easy to understand 57% 55% 50% 56% % of customers who were helped to solve their issues using KB 73% 79% 86% 71% % of customers who are satisfied with KB US English Permanent Spanish Permanent Spanish Pilot Japanese Pilot
- 29. V. Optimizing MT Quality
- 30. Steps to Optimal Quality . Increasing quality Controlled Language . Optimizing MT Quality Continuous Training Post-Editing Translation Memory Machine Translation
- 31. Dictionary Preparation . Data mining for terminology Bilingual files, TMs, glossaries, product names, GUI, style guides User dictionaries integrated into MT engine Verification, coding of terminology . Optimizing MT Quality
- 32. MT Virtuous Circle . Train MT engine on terminology for domain & client Linguistic Team Leader Process files (TM & MT) MT Engineer Correct output / Identify improvements Post-editors . Optimizing MT Quality
- 33. Full Post-Editing ▪ Publishable quality ▪ Terminology generally managed by MT engine ▪ Human effort to make more fluent sentences ▪ Quality metrics: same as with a traditional translation Light Post-Editing ▪ Understandable quality ▪ Terminology generally managed by MT engine ▪ May contain stylistic errors, awkward sentences ▪ Quality metrics: Can it be understood without reading the original? Can customers make informed purchasing decisions? Does it increase access to information? Does this customer support page solve clients’ problems? Does it decrease customer support calls/emails? . Optimizing MT Quality
- 34. ▪ Help/documentation, GUI, e-learning courseware ▪ Technical documentation, procedures , engineering reports, assistance/help desk ▪ Online knowledge bases, FAQs, user-generated content, real-time translation, streaming news, continuous updates & fixes, alerts, real time translation, streaming news, intranet ▪ E-mail, sms, customer support, company intranets, enterprise information ▪ Intelligence/Analytics (call centre logs, e-discovery, patents), search, CRM Suitable Content for MT . . Optimizing MT Quality
- 35. VI. Lexcelera’s MT Services
- 36. ▪ Established in 1986 ▪ Entities in Paris and London ▪ More than 2 decades of experience in the translation and localization industry ▪ First company in France ISO 9001 certified for quality ▪ Founder of Translators without Borders ▪ Pioneer in deploying optimized MT for enterprises Lexcelera . . Lexcelera’s MT services
- 37. ▪ MT Consulting ▪ Selection and training of MT Engines ▪ Integration ▪ Linguistic Preparation ▪ MT Implementation and Maintenance ▪ Terminology Management ▪ Post-Editing (Full and Light) ▪ Hosted MT ▪ … and more We provide: . Lexcelera’s MT services
- 38. ▪ Choose suitable content to translate VII. Keys to MT Success ▪ Choose the right system for your language pairs ▪ Know the limits of MT ▪ Invest in customizing and train fully before launch ▪ Author with MT in mind ▪ Get buy-in from your teams ▪ Provide clean data, memories, glossaries, etc. ▪ Integrate with your workflow ▪ Continually improve the engine
Editor's Notes
- #26: .