A first look at WikiCrowd

I have quite enjoyed the odd contribution to an app by Google called Crowdsource. You can find it either on the web, or also as an app.

Crowdsource allows people to throw data at Google in controlled ways to add to the massive pile of data that Google uses to improve its services and at the end of the day beat its competition.

It does this by providing a collection of micro contribution tasks in a marginally gamified way, similar to how Google Maps contributions get you Local Guide points etc. In Crowdsource you get a contribution count, a level, and a metric for agreements.

While I enjoy making the odd contribution when bored out of my mind and enjoy looking at the new challenges (currently at 2625 contributions), I always think that data like this should just be going out into the world under a free licence to benefit everyone.

So finally, introducing WikiCrowd, an interface, and soon to be app, that I developed over the new year period.

WikiCrowd Overview

WikiCrowd is hosted on toolforge and can be found at https://wikicrowd.toolforge.org/ (Source code on Github)

In order to contribute, you need some knowledge of the world, a Wikimedia account and that’s it!

I’m very aware there are an infinite number of things similar to this in the world, even within the Wikimedia space, but this is my take, with a few small existing differences, and some that are in the works.

The tasks that are presented are intended to be achievable by anyone, with very little context needed about what Wikipedia / Wikimedia is, how structured data works there, and what is actually happening behind the scenes.

There is currently a very heavy bias towards image depictions, as it turns out that Wikimedia Commons categories are a fairly good starting point for easy data sets, and people enjoy looking at images.

When selecting a category, you get presented with repeated questions, with a simple yes, no, maybe option.

Currently, if you hit “Yes”, and edit under your account will happen shortly after submission. “No” and “Maybe” are recorded, but for now nothing else will happen.

The idea here is that people like simple tasks, focusing on a single thing to identify in a picture.

Other views exist, such as proposing Wikidata statements from lead text of Wikipedia articles, which has been taken from the Wikidata Game by Magnus.

At the time of writing this post, 25 people have tried the tool out in the 5 days since it was put online. There are around 10k questions currently in the system (being expanded daily), and 5.3k of them are already answered leading to 3.7k edits on Wikimedia Commons and Wikidata.

The future

The easiest thing to see in the future is the continued expansion of the categories used for image depicts questions. If you have any ideas leave a comment on this post or reach out to me (perhaps on Twitter).

Really I want this to be a native phone app, which will bring offline access, faster access, and have it a little closer to folks fingertips. I needed a backend API before getting there, and accidentally created a preliminary UI in the process. I’m currently working on a phone app using flutter.dev.

Currently, all “Yes” responses immediately result in edits, but I want to introduce a concept of agreement. In most cases, this would be 2 or 3 yes responses = 1 edit, which should increase accuracy and avoid mistakes by fast fingers.

Likewise, all “No” or “Maybe” responses currently don’t get shown to anyone a second time. I’d like to change that for the same reason.

I feel that there is some value in the negative responses being public somehow. I’m sure when training machine learning models people would also like the negatives, but I’m not sure how I may expose those yet.