We wanted to do something related to natural language processing. So we found this Kaggle competition about classifying texts based on their toxicity and decided to give it a try.

What it does

It reads comments through a web interface and it reports multiple categories of possible hate-speech

How we built it

We used TensorFlow for the back-end and python/html for the front end. The machine learning model is creates embeddings for the most common words. Then, these embeddings in one sentence are combined into one which is then passed through a feed-forwards network to compute the result.

Challenges we ran into

Using google cloud platform (authentication), Firebase, and flask. The machine learning task using TensorFlow also got much more complicated, especially after not being able to figure out how to use sparse tensors.

Accomplishments that we're proud of

Building a natural language processing, and learning as much as we did in such short amount time. Finishing the hackathon with something that worked.

What we learned

We got a lot of experience with the back-end and we also learned how to use platforms/engines such as Google Cloud for Machine Learning. We also increased our expertise in Machine Learning with TensorFlow

What's next for Soft Spoken

Apply the algorithm to a wider variety of things such as Twitter, Facebook, YouTube, etc. Possibly create a chrome extension that would allow the user to filter hateful or inappropriate comments

Built With

Share this project: