C. Scott Ananian
| Tags | Censorship, Infrastructure, Languages, Machine Learning, Machine Translation, Translation, User Experience |
|---|---|
| Primary Session | Next Steps for Languages and Cross Project Collaboration |
| Secondary Sessions | Advancing the Contributor Experience |
'One World, One Wiki!' Instead of today's many siloed wikis, separated by language and project, our goal should be to re-establish a unified community of collaborators. We will still respect language and cultural differences - there will still be English, German, Hebrew, Arabic, etc. Wikipedias; they will disagree at times - but instead of separate domains, we'll embrace a single user experience with integrated navigation between projects and languages and the possibility of split screen views aligning related content. On a single page we can work on articles in different languages, or simultaneously edit textbook content and encyclopedia articles. Via machine translation we can facilitate conversations and collaborations spanning languages and projects, without forcing a single culture or perspective.
Machine translation plays a key role in removing these barriers and enabling new content and collaborators. We should invest in our own engineers and infrastructure supporting machine translation, especially between minority languages and script variants. Our editing community will continually improve our training data and translation engines, both by explicitly authoring parallel texts (as with the Content Translation Tool) and by micro-contributions such as clicking yes/no on a proposed translation or pair of parallel texts ('bandit learning'). Using 'zero-shot translation' models, our training data from 'big' wikis can improve the translation of 'small' wikis. Every contribution further improves the ability of our tools to make additional articles from other languages available.
A translation suggestion tool will suggest an edit in one language whenever an edit is made to a parallel text in another language. The correspondences can be manually created (for example, via the Content Translation Tool), but our translation engine can also automatically search for and score potential new correspondences, or prune old entries when the translation has drifted. Again, each new correspondence trains the engine and improves its ability to suggest further correspondences and edits.
Red-links and stubs are replaced with article text from one of the user's preferred fallback languages, perhaps split-screened with a machine translation into the user's primary language. This will keep 'small' language wikis sticky, and prevent readers from getting into the habit of searching in a 'big' language first.
We should build clusters specifically for training translation (and other) deep learning models. As a supplement to our relationships with statistical translation tools Moses and Apertium, we should partner with the OpenNMT project for modern neural machine translation research. We should investigate whether machine translation can replace LanguageConverter, our script conversion tool; conversely, our editing fluency in ANY language pair should approach what LanguageConverter provides for its supported languages.
By embracing unity between projects and erasing barriers between languages, we encourage the flow of diverse content from minority languages around the world into all of our wikis, as well as improving the availability of all of our content into indigenous languages. Language tools route around cultural or governmental censorship: by putting parallel texts and translations in the forefront of our UX we expose our differences and challenge preconceptions, learning from each other.
