About
I originally set out to make a website to illustrate the potential harm of social media.
My biggest challenge was finding a list of hate comments that were general enough to affect
almost anyone reading it. A lot of comments you find in data sets are the logs of chatbots or are found in
papers that are very specific and sometimes don’t even apply to humans (especially in the case of chatbots).
From there I decided it may be best to use an API, but it was difficult to make most text-generation APIs
generate mean comments (for good reason). The best I could do was
this API
that was fine-tuned using Republican tweets. It doesn’t always do a great job but there have definitely been
some stand-out moments with what it generates, for example:
“You are so stupid, you must be so ignorant because we
live in the land of the free!! What about the 70th amendment which allows a man to have 2 wives at once? That was
about equal opportunity for all!! I”
From there my direction changed and I decided upon a more content-moderation-themed approach.
When combined with a basic sentiment API I could both generate (mostly) negative tweets and decide
how negative they appeared to be. This became the framework for the current version of this website,
which aims to make the viewer think about how comments are moderated on massive sites, and what keywords
or general red flags an API may be looking for. There are times when a very negative comment receives a
low score due to the language that is used, and times when a relatively positive comment receives a very
high negativity rating. The sentiment API is not entirely representative of what a moderation API would be
looking for, as they are also looking for spam comments and bots in most cases, but it is a basic case of
how automated moderation works.