Hi everyone, hope y’all are doing fine. I’m working on a summarization problem where rather than simply summarizing a document, I have to create multiple smaller summaries, which when concatenated one after the other create a coherent structure, instead of simply having multiple self contained summaries of the document.
Haha, it would be close to that. But I expect it to be somewhat smaller in length, and with a different wording naturally. Sorry if it sounds like a weird problem but my employer asked me to probe more into how feasible it is.
One way would be to summarize different parts of the text individually. For example, summarize the intro, then summarize the body and finally the conclusion. Perhaps even let them feed off each other to maintain some sort literary flow. It might not be the most cost efficient approach, but I believe it would work.
Otherwise, if you just want to go off one document directly, you can maybe ask GPT-3 to produce those same results in a creative manner that only uses one prompt.
I will have to split the documents nonetheless, since most of them are longer than GPT-3’s context length. I was thinking maybe splitting into paragraphs of 10 lines each so that each query is about 500 tokens.