I've been using infranodus to analyse clusters of publication abstracts. In particular I've been trying out the GPT3 tools 'Paraphrase graph' and 'Essential and summary'. I upload all the abstracts in csv format with identifiers.
However, I've found that the summary produced (and, to a less extent, the topics identified) are sensitive to the order of the abstracts. If I reorder them (e.g. alphabetically) I get a very different summary. This is particularly the case for the 'Essential and summary' function., which tends to focus in on the first abstract in the sequence. Is there a way to solve this?
Yes, in fact the summarization part is based on GPT-3 AI, so it'll always generate something new. I'd recommend to run a few iterations every time. What you come up with will never be definitive, rather, it's good for ideating.
Particularly the Essential and Summary function makes use of the actual statements identified as key to the discourse, so that's why it depends on the order of them.
Please sign in to leave a comment.