-
Our hand built, grass fed logo
Inspiration
Our combined love of film and interest in neural networks made this project a excellent candidate for this event. We wanted to craft an item that provides a useful service to new and experienced writers alike.
What it does
Using famous movie scripts as training data tied to their entries in TVTropes.org, we used a neural network to find similar tropes in new scripts passed to it. We used a training data set of 98 movies, 7,201 unique tropes found on TVTropes.org
How we built it
We built a neural network using Python and TensorFlow, used Selenium to scrape tropes and scripts off of TVTropes.org and IMSDB.com respectively, and created a web front end using Flask, JavaScript, and HTML.
Challenges we ran into
The neural network was incredibly difficult to get operating correctly, and caused us to change our tactic part way through. Web scraping proved difficult due to inconsistent web design which forced us to be adept at error handling. The web front end was also difficult due to lack of experience.
Accomplishments that we're proud of
We built a complete working product that accurately finds tropes in scripts that are not in the training set. While these may not agree with TVTropes.org on every front, they seem to exemplify aspects of the films that were not mentioned by our source material.
What we learned
We learned how to construct neural networks for web use, as well as effectively web scrape and present results dynamically.
What's next for AutoTomato
We plan to further expand the training data set and refine the use of what we dub "Kettani valence" in order to determine the value of certain tropes for training purpose as opposed to others.
Log in or sign up for Devpost to join the conversation.