Determining the unit of analysis for the machine learning models

For the [chrome extension](https://github.com/Call-for-Code-for-Racial-Justice/TakeTwo-Marker-ChromeExtension), we need to decide what is the unit or type of selections trusted contributors can select when identifying racially biased content. This will be used to train the ML models and is important to help users understand why content is potentially racially biased and to offer up alternatives.

Considerations
* If content is considered racially biased, context will be a factor. How can we store more data around the selected word / phrase to provide more context for the ML models?
* At what granularity should the trusted contributors be expected to flag content? (Categorically vs binary?)
* Consider the input as a paragraph/group of text and all kind of tokenizations can be used to finally preprocess