blkr

blkr is a deep NLP algorithm designed to detect possibly offensive or annoying text on a webpage and remove it though a chrome extension. It is trained on a dataset of >500,000 tweets.

Comment

Inspiration

Twitter makes me angry but there are still some people that I would like to see updates from. Their platform is designed to provide controversial posts to keep you interested but it is not useful or enjoyable in the long term.

What it does

It uses an NLP algorithm built in TensorFlow to analyse tweets scraped from the webpage. The network classifies the tweet into several categories such as hate speech, political material, covid news and financial news. The posts are deleted if these predictions go over a threshold.

Challenges we ran into

Using pretrained NLP models can be tricky to implement and are overly complicated for a simple task. Therefore, I wrote the algorithm from scratch.

I also spent a long time attempting to get it to train on a TPU which did not work for reasons only known to google.

I also found that I ran out of google cloud GPU time half way through and had to switch to Kaggle which uses a different version of TensorFlow.

The chrome extension was (still is) horrible. I don't understand why it has to be so complicated. It might not work we will see.

Accomplishments that we're proud of

Achieving very good perfomance with a self designed neural network. The algorithm scored over a 92% accuracy.

What we learned

Not to use chrome extensions.

What's next for blkr

Generalising for more social medias.

Built With

Share this project:

Updates