TTTC's "detached island cluster" issue

A discussion of the phenomenon of strange clusters in places that fly like this in Talk to the City.

Observation Fact: There have been a number of duplicate examples given by fewshot to the outlying island clusters.

Guess:.

• Sometimes returns examples of fewshot instead of [] for input data that cannot be successfully opinion-extracted
• Maybe UMAP behaves as "there are clusters with high density" when many data points with almost the same content overlap.
  • I think the resulting increase in spatial resolution of the surrounding area gives the appearance of these remote islets.

Countermeasures:.

• If you show it to users as it is, most users will react, "What is this remote islet?" and will be the first to see it.
• Separate islets are "opinion data incorrectly extracted due to inadequate instructions to LLMs" and should be removed as a visualization of opinion.
• The data is the effect of "data that does not extract opinions well," so we will add a newshot to deal with that data.
  • Specifically, click on a point in this cluster to check the input data to the LLM, and if there is no opinion to be extracted, [], otherwise, add that opinion as a reply in the prompt fewshot and start over from the extraction phase
  • You don't have to do the whole cluster, just add one or two cases as a realization to get rid of it.
    • Maybe adding [] to the fewshot moves the boundary of "to [] or not".

As an aside, it would be better to UNIQUE it right after EXTRACTION in the first place...

This page is auto-translated from /nishio/TTTCの「離れ小島クラスタ」問題 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.