Editing Research/new digital reading experiences (section)

=== (Re)Interpreting content with Large Language Models ===

As mentioned earlier, we decided to emphasize parts of the text based on the words’ difficulty level. However, doing this manually can take way too much time. An example of such an intervention could be to find all mentions of colours in a text and then give all these words the colour they are refer to.

In this example of marking mentions of colour, we could quite easily write a script that would find all literal occurrences in the source text of a predefined list of colors and mark them. This sort of intervention has been done before and is not particularly exciting. What we want to try instead is to find many more ''abstract'' or ''conceptual'' elements in the text and mark those. An example of such an intervention would be to mark all emotionally charged parts of a text and design them in a way that matches that emotion. This is impossible to do with a classic script, and a lot of work to do by hand.

However, LLM’s (Large Language Models, e.g. ChatGPT) appear to be a great solution for this, so we decided to try out how we can utilize it. We are using ChatGPT 3.5, since at the time this was the cutting-edge tool.

<span id="process"></span>