We want to help people with psychological disorders that impair the ability to pick up social cues (i.e. schizophrenia, autism, social anxiety, ADHD, etc) and also English-language learners with an API that can determine whether a text statement is sarcastic or not. We decided on an API because we feel that this has multiple use cases, like a browser extension to check if forum posts are sarcastic, a phone app that checks how sarcastic your/others' texts are, or book reader app that highlights the sarcastic portions of a book.
What it does
Our API uses a Recurrent Neural Network combined with a statistical model using Bayesian statistics and word frequency to determine whether or not a sentence is sarcastic.
How we built it
We use Lua and the Torch library to train a neural network using sarcastic/non-sarcastic statements from a CSV file provided by the UCSC Natural Language and Dialogue Systems. After that, we made a REST API on a Ruby web server.
Challenges we ran into
- Time and Processing Power - If we had weeks and better computers, we could train our neural network with more data sets and become more accurate with our results.
Accomplishments that we're proud of
- Our first time making a neural network, and using the Torch library.
- Achieving a better success rate with determining sarcasm than we expected.
What we learned
- How to design and build a neural network
What's next for SassMaster
- Reading context around statements to better determine sarcasm.
- Training with more data sets to improve accuracy.