TextPositive

Inspiration

The Internet is a notoriously toxic environment. It is so easy to be mean or say offensive things either intentionally in rage or by accident. We made TextPositive in the hopes of trying to make the online community more aware. The ease of access for the internet especially in younger audiences and the rise of cyberbullying makes this issue all the more clear.

TextPositive is not a tool to censor the internet. We believe that the internet is as successful as it is because of free speech and without it, even the hate, it would not be so central in our lives today.

What it does

TextPositive is a Chrome Extension which reads the currently active element that is being typed in and tries to find negative sentiments in the text being typed. We use a deep learning recurrent neural network to find negative sentiments in the users text and suggest alternatives in a spell check manner.

How we built it

We used PaddlePaddle, a deep learning framework by Baidu, to train and deploy our models. We trained and deployed our models using a provided Amazon EC2 instance. Our website and Chrome Extension were built using HTML and Javascript.

Challenges we ran into

We ran into many challenges during our short period of time. Firstly, it was difficult training large models which take hours to train and also being able to fine tune these models with limited time to be deployed. Furthermore, it was difficult learning a new API and ecosystem for deep learning.

We also ran into the problem of modern web design. Scaling our product to be able to handle each website's different system of entering text was tedious and a work in progress.

Finally, we tried many techniques in PaddlePaddle for text generation but ultimately because of time for training settled on a much simpler approach that worked better in practice.

Accomplishments that we're proud of

We are proud that we were able to learn PaddlePaddle and Docker in such a short period of time, and were able to successfully train and deploy our models. We are also proud of building a highly usable and eye pleasing website and Chrome Extension for our product.

Most of all we are proud of the fact that we have created a product that has the potential of creating a better online community.

What we learned

We learned the difficulties of starting a project in a new Deep Learning Platform, PaddlePaddle, and having to learn the APIs, but also in the short period of time train and deploy our models. We also gained extensive knowledge of the Docker ecosystem, and figured out how to use Amazon EC2 instances to train and deploy models. We also learned how difficult it can be to create user friendly interfaces for websites.

What's next for TextPositive

TextPositive hopes to get better a negative sentiment detection and suggesting positive alternatives. We hope to be able to officially deploy on the Chrome Extension store and be able to help real people avoid negative commentary and create a more positive online environment.