Topic on Talk:Wikimedia Developer Summit/2017/Topic ideas

Figuring out how our current resource allocation might be shortsighted, what trends are we missing out on might be a good topic for the summit. My bet for that would be artificial intelligence (in the wide sense, including natural language processing, machine learning etc). User interfaces are getting more and more clever: they understand what you want without requiring you to become an expert in their syntax; they can guess what you will want next; they can relieve you from repetitive tasks by learning how to do it and let you focus on tasks with more added value. Your phone understands what you ask from it; the search engine you use tailors its answers to your past interests; soon your car will drive itself.

When the shift to mobile happened, Wikimedia was late to recognize its importance; out mobile interface was significantly delayed compared to other top sites, and had to be monkey-patched on top of a system that was not built with such a use case in mind. There is a danger that we repeat that history with AI. There are various initiatives that already involve or might benefit from AI tools (NLP in Discovery, article feeds based on past visits in the apps; suggestions in the editor; ORES; and chapter and third party initiatives like Wikispeech), but they are disconnected and not really part of any organization-level strategy, and in some cases suffers from inequality of investment and benefit (where everyone is happy to use an API but no one helps with its development - it seems ridiculous, for example, that the Foundation does not employ a single developer to work on ORES, which is really the only plan we have for addressing editor decline).

I might be misjudging the trends (AI is well outside my area of expertise) but my impression is that this area is way too important to happen in an ad hoc, skunkworks way; there should be a unified strategy, more support from the users to the providers of the machine learning infrastructure, and exchange of knowledge between Wikimedia developers working in a similar are, and between Wikimedia developers and external experts. WikiDev17 might be a good place to work towards that.

+1

I think that LZia and Ewulczyn will have a lot to contribute too.

Looks reasonable to me. So, +1.

Supporting this. Partial automatisation of image categorisation (which is basically classification) using deep learning/convolutional neural networks seems like one of the next logical steps in this area to me.

In the m:ORES project, we're currently predicting edit quality (vandalism? good-faith?) and article quality (Stub, GA, FA? and the equiv for russian and french). We're working on building models for predicting the type of change made in an edit (e.g. reformatting the current information or adding new information?) and detecting personal attacks in talk posts (based on work done in m:R:Detox).

A lot of AI relies on human-labeled data. So, one of our major projects is m:Wiki labels. Getting the resources to turn that into a proper mediawiki extension has been difficult. It seems that, if we can come together and discuss the value of AI and where we'd like to see support, finding resources for these types of technologies that all AIs use will be easier.  :)

Alright, so it seems that this proposal is still making it as main topic based on de facto consensus. Works for me.

We still need to fine tune the title of this main topic, just as we have done with the other ones. "AI" is too generic and doesn't provide any direction. What about AI?

See the other main topics at Wikimedia Developer Summit/2017/Program. This one needs to fit there.

I propose: "Leveraging AI to help build and navigate content"

Spelling out "AI", what about "Artificial Intelligence to build and navigate content". Although I am not sure what "navigate content here means".

"Artificial Intelligence to build and curate content"?

Can you be more specific about what you think is ambiguous about "navigate content"?

I didn't say "navigate content" was ambiguous. I said that I wasn't sure what it meant. After thinking more, I believe it might refer to AI helping readers to find interesting information based on their actions? I got to this conclusion because I am aware of experiments like creating related pages automatically or recommending articles for translation. Maybe it is only me getting confused about this "navigate content", but if I can be considered a benchmark, then many more people are going to be confused by this expression.

It seems you are interpreting "navigate content" in the way I hoped. :)

You've given a specific example of an AI that might recommend content based on past activity. We have AIs that generate new links based on people's browsing behavior. We have AIs that recommend articles based on general trends. You might imagine an AI that helps you find the right commons images for your new article. Both building and navigating content suffer from en:information overload issues. AI can help.

All yours: Wikimedia Developer Summit/2017/Artificial Intelligence to build and navigate content.