how to keep post-editors engaged and prevent attrition. (jose sanchez, ebay)
TRANSCRIPT
ENGAGING POST-EDITORS
JOSÉ LUIS BONILLA SÁNCHEZOctober 14, 2015
ENGAGING POST-EDITORS 2
TOPICS
• BACKGROUND: MT AT EBAY
• THE CHALLENGES OF POST-EDITING (INDUSTRY AND EBAY)
• THE EBAY EXPERIENCE: PROBLEMS AND SOLUTIONS
• A LOOK AT THE FUTURE
• DISCUSSION
ENGAGING POST-EDITORS 3
MT AT EBAY
•Since 2013
•Home-grown, statistical (Moses) engines
•Covering 10 language pairs (FR IT ESES ESLA DE BRPT •RU ZH USEN UKEN)
•Content translated: listings (item titles and descriptions), keywords
ENGAGING POST-EDITORS 4
CHALLENGES OF POST-EDITING
ENGAGING POST-EDITORS 5
INDUSTRY CHALLLENGES
Post-Editing projects add extra complications to the regular L10n flow. For instance:
- Quality expectations can be unclear (definitions of light and full post-editing vary).
- No universal agreement on rates (Per hour? Percentage discount? Edit distance?).
- Every project (and engine) comes with its own difficulties: statistical vs rule-based engines, technical vs social content…
ENGAGING POST-EDITORS 6
EBAY CHALLENGES
At eBay, we deal with especially complex projects:
- 12k+ eBay categories, many with their own terminology
- User-generated content: Unpredictable quality, slang, non-standard acronyms and abbreviations…
- Very specific requirements: Our goal is not polished content, but content which can be understood and useful to train the engine
Our Solutions
2014
ENGAGING POST-EDITORS 9
Modular Guidelines…
Language-
specific
Content-specific
General
ENGAGING POST-EDITORS 10
…Structured to Facilitate Learning
General introduction
PE-Specific Instructions
Item Titles Languages
RU
BRPT
ESLA
FR
IT
ESES
DE
ZH
Item Descriptions
QueriesGeneral
Translation Instructions
ENGAGING POST-EDITORS 11
Recorded Trainings
The more specialized the training……the more important to preserve the information in recorded format so future post-editors can refer to it.
ENGAGING POST-EDITORS 12
Escalation as Needed
3rd fail triggers a call with the vendor. • Participants: Linguists, PjM, Quality Manager• Agenda:- Diagnosis- Vendor Action Plan- Feedback for Client
ENGAGING POST-EDITORS 13
2014 RESULTS (FAILS VS PASSES)
1 2 3 4 5 6 7 8 9 10 11 120
5
10
15
20
25
30
35
40
45
Reviews Pass Fail
ENGAGING POST-EDITORS 14
2015
ENGAGING POST-EDITORS 15
EARLY SAMPLE REVIEWS
In-progress review for most problematic language combinations:
First 2k words of the project, in the first project week
ENGAGING POST-EDITORS 16
ADDITIONAL REFERENCE MATERIAL
Providing vendors with MT translations from 2 engines (generic and customized) so they pick the best
Out of Vocabulary words are unknown to the system, and left in English, which often makes them easy to mistake for brand names. By automatically tagging OOV, we allow the vendor to focus on them.
eBay listing titles can be difficult to understand without context (images, descriptions).
Alternative MT translation
Tagging “suspect” terms (OOV)
Providing full context (HTML files)
ENGAGING POST-EDITORS 17
RESULTS - PRODUCTIVITY
Series10%5%
10%15%20%25%30%35%40%45%
Productivity Increase
Alternative MT Translation
Tagging "suspect" terms (OOV)
Providing full context
ENGAGING POST-EDITORS 18
RESULTS - QUALITY
1 2 3 4 5 6 7 8 9 10 11 120
5
10
15
20
25
30
35
40
45
Pass/Fail over Time - 2014
Reviews Pass Fail
1 2 3 4 5 6 7 8 9 10 11 120
10
20
30
40
50
60
70
Pass/Fail over Time- 2015
Reviews Pass Fail
THE FUTURE
Divider sub-headline goes here
ENGAGING POST-EDITORS 20
THE NEXT BREAKTHROUGHS
By predicting the quality of the MT output, we could:- Filter and send to our vendors the best (or worst) translations.- Map QE score to time spent, and use it to calculate more accurate initial rates.
Tools like iOmegaT or Matecat offer time tracking, edit distance analysis and even action recording – usable to:- Analyze post-editor behavior and identify
areas to help them improve.- Calculate accurate rates.
The future is for online tools – they will allow for more direct interaction (early sampling, continuous communication, centralized information repository), further integrating the post-editors with the team.
Quality Estimation
Behavior Tracking
Online Collaboration
ENGAGING POST-EDITORS 21
DISCUSSION TIME