How We Built the Most Multilingual Translation System
Wikipedia Translation
Santhosh Thottingal
Wikimedia Language Engineering
Internationalization & Unicode Conference 39
Knowledge is better served�if you can understand it
https://secure.flickr.com/photos/adam_jones/5793940771/
English Wikipedia article coverage
http://ddll.inf.tu-dresden.de/web/Wikidata/Maps-06-2015/en
German Wikipedia article coverage
http://ddll.inf.tu-dresden.de/web/Wikidata/Maps-06-2015/en
French Wikipedia article coverage
http://ddll.inf.tu-dresden.de/web/Wikidata/Maps-06-2015/en
Polish Wikipedia article coverage
http://ddll.inf.tu-dresden.de/web/Wikidata/Maps-06-2015/en
Chinese Wikipedia article coverage
http://ddll.inf.tu-dresden.de/web/Wikidata/Maps-06-2015/en
Content overlap between languages of Wikipedia is small
Based on the analysis of one month of all edits to the top 46 language editions of Wikipedia by Scott A. Hale, University of Oxford.
More info at http://arxiv.org/abs/1312.0976
Big opportunities for translation
English
German
49%
7,102 languages in the world(ethnologue)
Demand
287 Wikipedia
Demand
7,102 languages in the world(ethnologue)
50%+ Monolingual users
Demand
287 Wikipedia
7,102 languages in the world(ethnologue)
Next billion online users in 5 years
Demand
50%+ Monolingual users
287 Wikipedia
7,102 languages in the world(ethnologue)
6500 new articles per day
Supply
Supply
70K contributors per month
6500 new articles per day
14K new accounts per month
Supply
70K contributors per month
6500 new articles per day
3 years�For at least 40K article in every language
It takes...
3 years�For at least 40K article in every language
It takes...
12 years
To double the size of wikipedia
Potential users are active
Over 15% of users edit multiple language editions.
These multilingual users are more active (2.3 times) than their monolingual counterparts on average.
Multilingual users made 30% of all edits.
Based on the analysis of one month of all edits to the top 46 language editions of Wikipedia by Scott A. Hale, University of Oxford.
More info at http://arxiv.org/abs/1312.0976
Main language pairs
Content translation is a beta feature
Available in Wikipedia in all languages
Workflow
Discover article lacking translation
Create translation draft
Publish as new article
Follow the current process but avoiding manual steps
Translation view
Where translations are made
Entry points
Ways to make users aware of the tool
Translation is not new for Wikipedia
Translation is not new for Wikipedia
Translation is not new for Wikipedia
Integration
Integration to contribution workflow
Hover menu on Contributions link
Integration to contribution workflow
My contributions page
Integration
Gray interlanguage link
The translation dashboard helps you keep track of your translations, continue ongoing translations and find articles to translate
Article and language selector
The translation tool
Machine Translation
Machine Translation
Available for 45 language pairs
Misuse of MT?
Translation progress is measured.
Translation context
Current sentence is highlighted.�Content is segmented at sentence level
Formatting preserved
Many machine translation systems support only �plain text translation.
Automatic link adaptation
Links can be added, removed or edited
One click image adaptation
Article categories
References
Translations always happen �at the target language wiki
Find source articles
Translate & Publish
Translations are auto saved, �can be resumed any time later
Link to source revision as attribution
Articles have ContentTranslation tag
Results
28000 new articles created
A new article in every 7 minutes
Next
Article Recommendations
Free Licensed
Parallel corpora
Based on translations of wikipedia articles
@WhatToTranslate
mediawiki.org/wiki/CX