TAUS Quality Evaluation Summit - 28 May 2015, Dublin
Over the past year we have been discovering that concepts once deemed “future of the translation industry” have become our day-to-day reality: translated content quality requirements have evolved, following the need to cover a wider range of translation scenarios, publish through multiple channels, do it faster and retire the content almost instantly. So what is king: "common sense" or "style guides"? Have we entered the age of "new creativity"? What does it mean to be a translator out in the field these days? We'll be discussing our findings around developing and deploying a viable "publish fast, forget fast" solution for various flavors of dynamic content, ranging from the ultimate "transcreation-level perfect" to "just actionable". The aspects we found to be of utmost importance are a very specific translator's skillset, a flexible, yet controlled evaluation environment, and, finally, the learning curve among the consumers of the translation.
4. Ephemeral but Here to
Stay
ONCE : “future of the translation industry”
NOW: our day-to-day reality
• Wide range of translation scenarios
• Evolving quality requirements – from “transcreation –perfect” to “just
actionable”
• Multiple delivery channels
• Retiring the content almost instantly
“Common sense" or “Style Guides” or both?
The age of "new creativity"?
When do we post-edit?
What does the workforce look like?
5. Story 1: Lightest PE Possible
Details
• Use MT but the content needs to be curated by a human to remove
the most prominent MT errors
• Delivering online content is the client‘s business
• Service, travel and hospitality domain - broad
• User-generated content, huge volumes
Translation objective: post-edit raw MT enough to remove the most
obvious MT errors and increase fluency
Key Challenges
Defining post-editing requirements
Achieving the highest productivity possible by avoiding over-editing
Making the content appear “human“ at minimal cost
6. Story 2: Real-time Content
Delivery
Details
• Cheap realtime translation
• > 30 languages, including rare language pairs
• Technical domain
• Professionally authored content
Translation objective: gisting quality MT; identify main areas of traffic for future
investment; obtain in-domain data to fine tune systems
Key Challenges
Range of languages and sparsity of training data for many of the languages
Post-editing for MT engine fine-tuning – training data volume requirements
Measuring the content impact and success
Engaging community?
7. Story 3: PE for Machine
Consumption
Details
• Gisting quality raw MT + post-editing of MT for SMT corpus building
• All main language groups
• Extremely wide range of consumer goods
• User-generated; cryptic / abbreviated; elements of SEO
Translation objective: post-editing to build good quality in-domain corpus
for improving SMT corpus > improve raw MT output
Key Challenges
Identifying the right translation resources – professional or just internet-
savvy
Define post-editing requirements – style guide or common sense, or both?
Achieving the desired productivity gains
8. Story 4: Raw or Light PE for
Sales Enablement?
Details
• Gist translation of bid-related documentation
• Into-English translation direction
• Big volume legal/business documentation
• Well-authored but contains complex “legalese“ clauses
Translation objective: provide gist translation of critical legal documentation to win a
bid
Key Challenges
Selecting the best translation process / resources
9. Story 5: Actionable MT
Output
Details
• Real-time communication support
• English Germanic and Romance translation directions
• Content authored by field users and support personnel
Translation objective: provide cross-language real time support to
European distributors
Key Challenges
Achieving the balance between Adequacy and Fluency when
introducing training data
Delivering usable and actionable instructions
Overcoming SMT shortcomings when handling negative/positive
statements
Educating „the casual evaluator“ and getting their acceptance
10. Findings
TRANSLATION WORKFLOW
Technical setup often very simple, but tailored to the request
Initial request may be for raw MT, PE requirement becomes necessary later
TRANSLATION SUPPLY CHAIN
Need for flat hierarchies, direct interaction, not suitable for all structures /
professional translation providers
Selection through motivation for unconventional translation methods
Shift focus to content and purpose, away from ‘consistency with legacy
translations, style guides & unvalidated glossaries‘ type setups
QUALITY EVALUATION
Less needed to PASS / FAIL a final translation, but crucial for aligning
expectations, defining MT engine retraining needs, establishing PE
requirements