Get an emotional snapshot of your texting style and contribute to AI research

Let’s run a data experiment … together. Introducing the Snips Emotional Snapshot.

Joseph Dureau
Snips Blog
6 min readSep 13, 2016

--

TL;DR

  • We are working on an assistant that reacts to the content of your messages, whatever the platform (Whatsapp, Facebook Messenger, etc).
  • We are doing so without collecting any data from our users. Your data stays on your device!
  • We still need data for our AI research, so we create experiments where we explicitly ask for data contributions in exchange for something cool.
  • Emotional Snapshot is a MacOS application where YOU can contribute with your text messages, carefully anonymized and encrypted to ensure your privacy before any data is uploaded.
  • As suggested by its name, the Emotional Snapshot app will thank you with a visual guide to the emotions you express in your texting life!
  • Curious about other Data collection experiments at Snips?
How joyful are you in your text messages? Loving? Grumpy? Do you want to figure out? Try out the Snips Emotional Snapshot.

Welcome into our lab 😜! At Snips, we are striving to build an assistant that will make technology disappear. What does that mean? It means we want you to get back to your life, and let technology handle the rest. Because Snips builds AI assistants that are private by design, we never get access to your data. We rely on intrepid explorers like you to contribute! We are asking YOU to contribute your anonymized chat messages to science to help us create better AI that is private by design. And as a special thank you for contributing, you will get in the application a visual guide to the emotions you express in your texting life!

Concretely, we’re working on various AI experiments, including a bubble that pops in situations in which you may need it (see this post for more details). It can be triggered by your context, or by the content of your messages. For example, when we understand from your messages that you’re going to meet someone in a place you’re not used to going to, our bubble can help you out with the corresponding directions.

Apple and Google are experimenting with this kind of ubiquitous assistant technology, but they are only ubiquitous in the realm of iMessages and Allo respectively, their proprietary messaging solutions. So let’s make the experiment together: let’s build a technology that works everywhere (meaning across all your favorite messaging platforms, Whatsapp, Facebook Messenger, etc), and would come at a zero privacy cost for users!

Building an AI that understands the way we talk

The core challenge here is to recognize the things (dates, locations, emotions, events, etc) we mention in our messages. For a start, we are mainly looking for references at locations, which is a fundamentally hard problem: it is harder than identifying references to people, for example, because there is a much wider variety of names for restaurants, bars, museums, shops, etc, than there are for people. Identifying locations is also very hard because the way we write our messages is so human: rare capitalizations, typos, exotic grammar and expressions…

However, the fact that we communicate in a uniquely human way isn’t something that can’t be handled, as long as we get the right data. We’ve scraped every last penny of our team’s text messages, and we’ve even processed every subtitle of every episode of about 300 tv shows. By relying on series like Girls, The Game, Friday Night Lights, West Wing, Black-ish, The Wire, Treme, or Jane the Virgin, we get an incredibly diverse grasp of the variety of ways we express ourselves.

We have made very good progress with this material, but in order to give users the best possible, most reliable AI product, we actually need real text messaging data. More than what we can collect among ourselves. And because the applications we create comes with zero privacy leakage, we are transparently asking you for your help. Will you donate your texts to science? Download the Snips Emotional Snapshot, and let’s run an experiment together!

One small slice of data for you, one giant piece of knowledge for AI-kind

Of course now you’re probably thinking twice (if not more) about agreeing to contribute your data. That is why we’ve taken every possible precaution we could take to protect your privacy.

To make this anonymous texting data donation possible, we have built the Snips Emotional Snapshot: an OSX app that you can download here. For the moment, it only works for owners of an iPhone and an Apple computer. This application is going to connect to iTunes, and then fetch your iMessage and Whatsapp messages from you iPhone. Here is what is going to happen, before any data leaves your computer:

  • You’ll be able to filter out any message or conversation you deem too sensitive.
  • We carefully remove any mention to phone number, credit card or social security numbers. That’s not the kind of value we see in your data. ;)
  • We only consider the messages you sent.
  • We round up the timestamps of your messages.
  • We delete any ID that links you with your messages (including our database logs).
  • We use public key encryption in the app to make sure only a selected group of data scientists at Snips have access to the data.

We even tried stylometric fuzziers, to make sure no one could identify someone’s message from their writing style, but we found that it destroyed too much of the meaning of the messages, making them useless for research.

What’s going to happen with my data?

As mentioned, it’s going to be encrypted, and mixed up with other people’s messages on one of our servers. Our current objective is to better understand what the words are that you use to to express that you’re interested in a location, so we’re going to use an algorithm to filter out all of the around 90% of messages that obviously don’t refer to any location. What comes through will be sent to external workers on Mechanical Turk, that will see the messages, one at a time, and will indicate for each message if the author is expressing the intention to go to a given location. We will make sure workers are correctly paid, and will not see more than a certain amount of messages. We will then use the results of this experiment to train our models to understand the human, conversational ways we talk, and to bring some magic to your phones.

Are you joyful? Loving? Grumpy? Get your Emotional Snapshot!

Because this experiment may take a few weeks, we wanted to give you a little something to express how thankful we are for your contribution. We’ll be giving everyone who donates a unique insight into the precise types of emotions they use when talking to friends. This insight comes in the form of a visualization we feel warmly about (and that we’re going to ship in an Android app coming soon). Using your own unique texting data, we can show you the emotions you are sharing with each of your contacts. Remember Inside out? This crazy crowd of emotions playing with your mind? Just use our app, and see how these emotions are playing with you when you communicate with each of your colleagues, friends and family.

GIF via “The Birdcage”

Thanks again for reading this post, and for taking your Snips Emotional Snapshot! We will be keeping you posted about our progress on this experiment, and thanks to all of you, we’ll be making AI that understands us a little closer to a reality! If you liked the idea of crowdsourcing user data instead of forcing user consent, have a look at our other data collection experiments! Who said AI wasn’t a warm collective adventure? ;)

If you enjoyed this article, it would really help if you hit recommend below :)

Curious for more? Have a look at all of our other data collection experiments!

Follow us on twitter @jodureau & @snips

If you want to work on AI + Privacy, check our jobs page!

--

--