Very cool and congrats on the launch. I see your demo is on android, are you doing anything with iOS / CoreML? I put up a similar project for OTA updates on iOS using CoreML to hot swap out device side models https://github.com/rkirkendall/MLEdgeDeploy
Thanks for looking at the repo. A couple of things to keep in mind.
- This list is only a starting point I pulled from a white paper. I agree that it does need to expand, hence the ongoing and open source nature of the project.
- If you have a suggestion for how this could be done with machine learning, please feel free to submit a pull request. Just keep in mind that the project only aims to catch language that can be reasonably interpreted as self-threatening.
Lastly, if you have a problem with how the project is implemented PRs are more productive than comments. It is far from perfect, but I believe it is still better than nothing at all.
Good question. Part of the ongoing nature of this project is to expand our phrase detections. I pulled the original phrase list from a white paper written by BYU last year because I figured that would be a good starting point. There is room for vast improvement though.
If the last few posts leading up to the flagged post are classified as predominately negative (currently looking at the last 3 tweets before the flagged tweet), then we send the notification. The intuition is that if a person is comfortable enough with social to post seriously suicidal content, he or she has probably already made some preceding negative remarks.
Hey thanks for the feedback. I understand your point on the topic, but I would like to point out that CheckUp isn't trying to 'solve' suicide in this manner. The philosophy behind the project is that you, at some level, care about the people you are socially connected to, and you may care if they are contemplating something very bad.
The goal is to create a service that will prioritize self-threatening posts from people you care about above the usual noise of your social network. Publicly posted cries for help can indicate very serious intent and we just want to make sure they don't go unnoticed.
Hey! Thanks for the heads up. We picked up more traction than I had anticipated, but we should be back online now! And some people may be interested in the Go repo ;)
Hey thanks for the feedback! I will try to update the site to clarify the language a little more –– what you described actually is how it works. The user that signs up is the concerned party that we would notify. The app watches all tweets on that user's home timeline. By "home timeline", I was referring to all tweets posted by the user and everyone followed by that users (so, your Twitter feed). That's how it's described in Twitter's API docs, but I can see where that may be a point of confusion.
"Because, as Y Combinator's Paul Graham puts it, you can't ride a Segway without looking like a "smug dork." And people generally try to avoid looking like that. "
Thinking about Google Glass, I feel like the author may have picked the wrong new tech to compare to a Segway. I still wouldn't agree with it, but he would have a better argument.