• Problem: What problem are you facing that needs support to be resolved or mitigated?

Currently, the Tone Check model is able to support Tone Check and Revise Tone features in 17 languages: English, Spanish, French, Portuguese, Japanese, Arabic, German, Persian/Farsi, Italian, Polish, Romanian, Russian, Turkish, Chinese, Czech, Indonesian, Latvian. We would like to enable these features in as many languages as possible, and we need to validate that the model performs above a certain threshold of precision in all additional languages.

• [Optional] Possible solutions: What improvement(s) would most meaningfully help you with this problem? Feel free to suggest multiple ideas.

We should collect ~1000 samples of labeled data for each of the languages that we want to expand to, and test the model before releasing to these wikis.

• Enabled projects: Which specific user-facing features or experiments would be unblocked or meaningfully enabled (in terms of development ease, velocity, etc.) by solving this problem? Which teams are launching these features or experiments?

We want to finish rolling out the Tone Check and Revise Tone features in additional wikis.

• Urgency and importance: When are these features or experiments expected to launch? How essential is this infrastructure for unblocking development?

@ppelberg to add: Which wikis are priorities for the rollout of this feature? How urgent is this work, and what factors affect the timeline and urgency of this request?

• [Optional] Notes: Is there anything else you'd like to share?

Evaluating the model's performance on a per-wiki basis is not just about our ability to interpret the edit text in each of these languages (via the underlying BERT model) - it's also about the model's ability to apply wiki-specific rules and policies about what is or isn't considered biased language. Thus, wikis with more detailed policy documentation and more organized ways of tracking examples of biased content will be favored in this approach.