Initial Code Extraction Process

[erikaz1]

Hi Thomas, my name is Erika, I am a grad student at UChicago in the States. Your AcademiaOS paper and UI is fascinating and absolutely amazing work.

In the 2023 paper, you write, "First, the system creates initial codes from the raw documents. Initial codes are short text strings describing emergent themes, concepts, and patterns in the language of the raw document" (5). Do you have any idea how exactly GPT chooses its particular output strings/themes? This is an ongoing concern I have regarding Grounded Theory coding since without leading research questions or directions, a text can have a range of key, albeit tangential, themes. This further begs the question, how do we determine the cutoff for this long list of concepts?

I've independently run prompts to GPT turbo/40-mini asking it to explain its initial code selection process. It said it looks for repeated words/phrases, phrases with emphasis, and probably uses context clues to find main ideas. Some of the decisions could not be explained with specificity.

Thanks again for sharing your awesome research.