What if inappropriate comments could be flagged as you type them out?

Many of us know the terrible feeling of unintentionally saying something rude or insensitive. Some of us know that on the global scale of the internet, problems with miscommunication and misunderstanding are very serious. We believe that most people mean well on the internet, and our project is to bring awareness to enable people to carry out their best intentions.

What it does

Browser plug-in that identifies when the user's social media comment/post contains language that is similar to other comments/posts that have sparked conflict or elicited backlash

How we built it

  1. data processing pipeline that downloads an online database of Tweets, passes the Tweets through sentiment analysis software (Google and Microsoft APIs), and scores each Tweet based on the "consensus sentiment" among its replies and quotes. (All of this done in parallel across dozens of computing nodes on Google Cloud)

  2. hybrid RNN/CNN-based model that predicts how likely a post is to elicit a negative response.

  3. Chrome extension that uses the AI model to recommend users to reconsider their comments or posts if the model detects a very inflammatory post.

Challenges we ran into

Data is hard to obtain and process. Even though Twitter is technically open-source, each day of tweets is 1 GB (!) and not everyone has 1 TB free to spare (thankfully Google does). Big data means big power, but also big costs, smart optimizations, and effective parallelization across CPUs/GPUs.

Accomplishments that we're proud of

Tackling a very serious problem, providing a feasible solution that we can continue to develop, developing effective data-mining pipelines, using a lot of computing power (100+ CPU-hours)

What we learned

Truth is objective but hard to quantify. Emotions are subjective and still hard to quantify. Sentiment analysis is cool but social media (with its abbreviations, emojis, etc.) is nuanced

What's next for Trigger Warning

Users! Both end-users and websites that want to incorporate our tech/ideas natively

Share this project: