Content translation defines a set of limits to encourage users to review the initial automatic translations (and even prevent from publishing) when the user has not edited the contents enough.
For this system to work well we need to find a good balance about when and how to warn the user and making sure that it works well for different languages, types of content and machine translation quality levels.
Based on analysis, research and feedback received, we may need to polish different aspects. This Epic ticket compiles improvements in this area.
Calculation of content modifications:
Limits to identifying content as problematic:
- T245840: Adjust the system of limits to make it more predictable
- T215403: CX2: Don't apply unmodified content restrictions to translations published to the user namespace
- T331023: Limit the publication of fast unreviewed translations
Communicating the problem to the user:
- T285861: MT warnings showing too early as the paragraph loses focus
- T263116: Once the MT warning is shown for a paragraph update it as the user edits its contents
- T202544: CX2: Make "mark as resolved" feature work across sessions
Enabling feedback:
Evaluation:
Handling specific contents:
Language-specific adjustments:
- T276895: Link to community-specific translation guidelines
- T246383: Adjust the threshold for Chinese Wikipedia to prevent publishing when overall unmodified content is higher than 70%
- T245509: Adjust the threshold for Assamese to prevent publishing when overall unmodified content is higher than 70%
- T228971: Adjust the threshold for Indonesian to prevent publishing when overall unmodified content is higher than 70%
- T244769: Adjust the threshold for Telugu to prevent publishing when overall unmodified content is higher than 70%