Crowdsourcing a language, as a game
Indonesian-language machine learning had a plain bottleneck: not enough labelled data. The data could not be scraped. Indonesia holds hundreds of regional languages and a slang that shifts month to month, and only people carry that knowledge. The problem was how to get thousands of them to hand it over, when none of them had a reason to.
The fix treated this as a motivation problem, not a data problem. It borrowed the logic of human computation: take a task only people can do, and wrap it in something people want to do anyway. A CAPTCHA uses that structure to block. Here it was inverted into a game. Players translated and validated words between Bahasa Indonesia and regional languages, ranked on a public leaderboard by volume and accuracy.
The leaderboard did the work that pay or obligation usually does. It made a repetitive task feel worth winning, and cut a large research problem into small moves that were satisfying on their own. The result was a dataset built, for free, by a country’s worth of contributors.
Motivation does not have to be ideological to be real. The line between a gate and a game is mostly a matter of who keeps the score.