-
An example of our chatbot and model in action!
Team Number: 67 Team Name: Basis
To try this project, message +18106425397. Do note that there are only 10 dollars in Twilio credits left, so don't spam it too much :)
Inspiration
As a users of Instagram, we are often flooded with bots asking for inappropriate things, creeps in comments sections, and real predators attempting to lure our friends in their DMs. We decided to do something about it for this Hackathon, and create an AI model which can assign a risk score to messages.
What it does
Our project is comprised of systems - the model and the flask webserver. The flask webserver powers the SMS chatbot which you can interact with. The SMS chatbot also assigns each message of yours a risk score - which can easily be used to debug the risk scores and see if the AI is working.
How we built it
We first created a tensorflow model on how we want the model to work. We used an embedding layer to parse vocabulary as sentiment, and use an LSTM layer to analyze the entire message. This is then shifted to one node, which is the probability of a message containing predator behavior. To train our AI, we used the PAN2012 dataset.
Once our model was complete, we had an accuracy of about 83%. This was good, so we decided to create a chat bot which you could talk to and gave you a score based on your messages. We were initially going to do an app which scanned your messages, but sadly we didn't have enough time.
To do this we used Twilio's API and a flask webserver to parse messages, send them to our chatbot, and send the message to our AI. The chatbot was based on some existing code but heavily modified to fit our purpose.
Challenges we ran into
We ran into a lot of challenges. First of all, finding datasets for our idea was pretty tough. It wasn't until Saturday noon our request for PAN2012 dataset was approved, greatly slowing us down.
Because of these time constraints, we were unable to create what we initially wanted for the project, including a web demo and app to scan incoming messages. However, we still think our little chatbot is cool!
Along with this, the model isn't the most performant and could be optimized. Hosting the project was also a pain, as GCP doesn't allow you to use a GPU instance on trial credits. Luckily, we got around this by locally hosting it and using an ngrok tunnel (rather jank).
Accomplishments that we're proud of
We're really proud of what we were able to do. Create an AI to stop predators, something that directly affects thousands of children worldwide. We didn't think it was possible or that we would even have time to submit. This also helps the burden on moderators caused by increased messaging during the pandemic, as an AI can flag messages for them.
What we learned
We learned a lot about Artificial Intelligence and how Tensorflow works in general. This was our first time touching Python in a while, so it definitely added time constraints. We also learned to reduce the scope of our project so we can get it done in time.
What's next for Groomer Remover
We're going to create a mobile app which can scan incoming messages from apps like FaceBook messenger, Instagram, iMessage, and normal SMS for predatory behavior and let the user know by using our existing model.
Log in or sign up for Devpost to join the conversation.