Inspiration
Inspired by Themis, the goddess of knowledge/justice, we aimed to develop a machine learning model to understand potential biases in the news that we read.
What it does
- Users submit a news article URL.
- The application scrapes and processes the article text.
- Text is passed through two AI models trained to detect bias.
- A weighted algorithm combines the outputs into a single bias confidence score.
- Results are displayed neatly on the web interface.
TL;DR You submit URL --> Application process article and detects evidence of bias --> You get snippets of the article where bias is most prevalent and an overall bias score.
How we built it
We trained our own BERT model on an existing dataset of biased/non-biased articles, and we used an existing bias detection model. We calculated a final weighted average bias score using both models and outputted this score into our UI on the website.
Challenges we ran into
Training our own model took a lot of time and compute and there were a lot of limitations we needed to get past in order to get something usable. However, using another model in conjunction with ours helped alleviate that.
Accomplishments that we're proud of
Building our own ML model, finetuning it, training it, (and using it alongside a model found on Huggingface).
What we learned
How to build a ML model from scratch. We learned and applied a lot of the fundamentals of data science from some of our classes, internships, and other extracurriculars.
What's next for Themis
Moving it to a chrome extension to making it portable and reduce the amount of steps it takes for the user to detect bias. We want to convert our models to work in ONNX Runtime so that they can run in the browser.
Log in or sign up for Devpost to join the conversation.