4. The many uses of machine translation technology
Let Google and Microsoft run with it or invest in your own translation technology. Why build your own machine translation system if Google and Microsoft offer such a great machine translation service.
Machine translation technology is a force multiplier and a catalyst for innovation. Learn more about how effective use of MT opens many new services and markets.
Panelists: Stéphane Domisse (John Deere), Olga Beregovaya (Welocalize), Diego Bartolome (tauyou), Dragos Munteanu (SDL), Tony O’Dowd (KantanMT), Sanna Piha (Moravia), Irene O'Riordan (Microsoft)
Presentation on how to chat with PDF using ChatGPT code interpreter
Irene O'Riordan (Microsoft) at the Industry Leaders Forum 2015
1.
2. Raw-MT goals (Support.Office.com)
Raw-MT article published within 24 hours
MT first: use MT on all articles, human translate high value and high traffic, within 5-10 days
24 languages production approved
90%+ of support articles (10,000 per language, 100+ new/updated per month)
MT only for long tail, bottom 30% of traffic
Automatic translation (recycling + MT) for both text and art
Target ‘acceptable’ quality: >=50% Customer Sat (CSAT), or within 10% of Human translation CSAT
Articles below the threshold are not published
4. Quality model
• Raw-MT translation/content quality bar: ‘acceptable’
• MT needs to reach a minimum bar to be usable
• Thresholds are 2.5 /4 for human survey assessment, and 50% CSAT
• Recycling helps quality, factored into quality metric – if quality is not sufficient Recycling threshold will be
used
• Onboarding
• 36 main SKU languages have been evaluated for MT quality
• Format: surveys with LSPs, in country and volunteer native speakers
• Special onboarding project in iCMS for limited live publishing, for eval and BI/ratings
• Continuous assessment
• BI monitoring of ratings to ensure MT quality bar continues to be met
• Upgrading to HT
• High traffic MT assets will be upgraded to HT, to ensure optimal customer experience
• Functional/non-language specific quality bar
• Similar to HT topics + disclaimer check & translation coverage
Top business goals for raw-MT usage is turnaround and long tail
Scope:
Goal is to use for as many languages as possible, publishing in 39 intl languages, 36 languages evaluated (2 further in eval at the moment), 24 currently production ready.
Expectation is 90%+ of all support articles can be machine translated
Only articles are currently in scope, other content types like training, video, templates will be looked at in future. Marketing content is also excluded.
Automatic translation of art: working on solution for screenshots leveraging OCR, recycling against software, and MT fall-back, have had promising results.
Quality
Measured by user feedback ratings
Recycling used to improve quality by reuse, and to also gate quality. Recycling level is used to determine if articles should be published, for conditionally approved languages
Metrics
1.Speed: 24 h turnaround
Business goal 1: Reduce time to market: Use raw-MT to publish content quicker than human translation
Metric 1. Turnaround time, 24 hours from when an English article goes live, to when a translated article goes live, for support.com
2. Scope: Traffic & volume – 70% HT page-views
Business goal 2: Focus human translation on high value and high traffic content, use MT for long tail and to deliver more volume and added value.
Metric 2: Traffic: The top 70% of page views for each language where MT is enabled should hit human translated content. The bottom 30% is MTd (assumes 80/20 traffic distribution where small amount of content generates most traffic)
Metric 3: Volume: MT should be used as much as possible, subject to availability and quality and acceptability constraints.
3. Quality: acceptable
Business goal 3: Quality & Customer experience: published MT content, needs to be understandable, of acceptable quality, and not detract from overall customer satisfaction
Metric 4: BI ratings: Customer Satisfaction rating should be >=50%, or within 10% of HT or English
Metric 5: Human evaluation: average score 2.5 or above.
Disclaimer at top and bottom.
Top disclaimer contains link to US topic
No bilingual view, planned for future
MT quality evaluation for Office.support.com
Human evaluation on 4 point scale, standard survey format.
Per language: 10 articles, 5 reviewers from LSP + additional in market /native speakers for tier 1 and 3.
5 articles with no recycling, 5 articles with 50% recycling.
Result based on main question: was this topic helpful?