We created dontfowl.me after reading Susan Fowler's article. We think it is high time we become more inclusive in tech. However, there were no NLP models that could detect harassment accurately. Even the most effective model, n-gram has a high false positive rate. dontfowl.me is a unique API that uses a new model called LSF (lexicon-syntax feature) to detect harassment.
What it does
When you send a block to text to our APIs or NLP golang library, we tag all harassing comments. As a demo we even made a chrome extension that detects sexually harassing comments in your email even before you send them.
How we built it
We used Stanford's coreNLP libraries to do basic NLP such as parts of speech tagging and creating dependency graphs. Then we wrote golang scripts to take these trees and parse them into one huge tree which maps syntax as well as parts of words and then eliminates harassing phrases using LSF.
Challenges we ran into
Setting up servers was hard because of Azure's security rules and NLP libraries quirks on slow machines.
Accomplishments that we're proud of
Building an entire NLP model in 36 hours.
What we learned
Both of us were new to golang so we spent most of our time messing around with golang to improve our code constantly. We were also new to NLP and learned everything at the hackathon and even came up with a better NLP model than all existing ones.
What's next for DontFowl.me
Documentation, tests and efficiency improvements. Hopefully more devs use this to create better applications