Overview | Developer Guide | Artificial Intelligence | License and Credit
Try the extension for yourself, available on the Chrome Web Store!
Bias Monitor is a Chrome extension that detects the political bias of news articles. Using Natural Language Processing (NLP) a machine learning algorithm analyzes the context of the website content. The algorithm determines the political bias of the website, and the Chrome extension displays this information for the user to see.
The extension also remembers what your viewing habits are and what your own political bias is. Bias Monitor learns about the type of political content you are used to seeing and encourages viewing news sources of a conflicting perspective.
Please leave any complaints or suggestions in the Github Issues tab.
To contribute to this project, create a pull request against the
main branch. One of the owners of the project will review the
request and determine if it may be merged. See the
CONTRIBUTING.md file for further reading.
Download the repo and follow the
Getting Started
documentation on the Chrome Developers website.
When loading the unpacked chrome extension, specify the dist
folder. To develop on the project, you must use the dev
environment. The dev environment returns mock responses and does
not hit the API. Create a config.js file in the project
root directory with the following pattern:
// config.js
module.exports = {
environment: "dev",
api_url: "<your-api-url-endpoint-here>",
api_key: "<your-api-key-here>",
}Since neither the API URL nor the API key is public, they do not need to be specified and mock responses will be returned.
Bias Monitor uses a Machine Learning (ML) algorithm to determine the political lean of text input. More specifically, the area of ML this algorithm is concerned with is Natural Language Processing (NLP). To create the algorithm, labelled data was used. That is data that acts as input and correlates to a "bias value" indicating its political lean. By minimizing the difference between the actual bias value the ML algorithm produced and the expected value provided by the labelled data, the program can—in a sense, learn.
To train the AI, labelled data is required. Three different labelled datasets were used in the training process. The first dataset came from a pre-labelled collection of articles from AdFontesMedia, providing approximately 1200 entries, totalling about 1 million words. The data was labelled from -42 to 42, representing far left to far right respectively.
Other datasets were created by piping data through Media Bias Fact Check. The Media Bias Fact Check website analyzes data and gives a left or right political lean. About 10 thousand articles were labelled and trained on using this method.
A third data source was gathered from All The News, a Kaggle dataset consisting of about 100 thousand articles labelled based on political lean.
The API takes in text data, analyzes it, and returns a number in the domain of [-42, 42]; a more negative number represents further left politically, and a greater positive number corresponds to a far-right political lean.
The two main architectures used in finding the best model were Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT). LSTM is what lays underneath ULMFiT which was used to train one model. BERT is another neural network learning approach used in determining the best model. RoBERTa is a branch off of BERT and proved to provide the best results of all the ML models considered.
ReBERTa achieved the best Margin Absolute Error (MAE) of 6.04.
Thanks to the wonderful report by The Bipartisan Press and for providing an API endpoint for our Chrome extension project Bias Monitor.
Licensed under the MIT License.