prev Broad Listening from Text Format Data
- Hypothesis that customer value is the discovery of a “new angle” when the objective is not to visualize and show a scatter plot
mounting
- Instead of making it 2-dimensional with UMAP and then SpectralClustering, HDBSCAN it in a higher dimension.
- In this case, use a large number of clusters (30~100), discard sparse regions, and extract only dense clusters.
- When we did “Teaching AI the KJ Method”, we had the students extract pairs as if they were teaching humans, but this time, we will directly perform the “grouping” of the KJ method by extracting the dense regions in the high-dimensional space.
- Have LLM create a description of the extracted clusters (where Surprisingness Judgment is also performed).
- This is what the KJ method calls “making a table nameplate.”
- Describe the group first, then add a front cover that compresses that description.
- Then, based on the description of the cluster, check for “relatedness” with respect to the excluded data, and pick up
- This is a confluence of ape.
- Expanding associative mood and expansion of pertinent judgment increases the probability of new associative connection.
- This is a confluence of ape.
When the objective is to discover new viewpoints, “there were many opinions like this” is not important, so the direction is to “extract strong opinions,” “discover unexpected approaches,” and “take in a wide range of information related to that.
PS
- In the above process, I abandoned the creation of a scatter plot once, but then I realized that in the KJ method, too, you do the grouping and then the spatial arrangement, so you can just do the spatial arrangement after this.
This page is auto-translated from /nishio/新しい切り口を発見するためのブロードリスニング 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.