pConceptMap2025-09-08

memo

4 stages: (1) sectioning → (2) concept extraction → (3) relationship identification → (4) integration

Labels for extracted concepts

Identification of the relationship between them

---

Summary of experimental results:

📊 Generated concepts (5):

globalization - core

digital technology - core

democracy - core

plurality - core

collaboration - supplementary

🔗 Relationships between concepts (4):

globalization → digital_technology (prerequisite_of, 0.90)

democracy ← collaboration (part_of, 0.85)

digital_technology → plurality (example_of, 0.80)

democracy ↔ globalization (contrasts_with, 0.75)

---

I understood the approximate flow.

First, "keywords that seem important" are extracted from the given document,

Then give that keyword list and a set of documents to extract the "relationships".

cost consideration

Now gpt-4o-mini

The fee for developers is 15 cents per million input tokens and 60 cents per million output tokens (equivalent to about 2500 pages in a standard book).... . ChatGPT will allow Free, Plus, and Team users to access GPT-4o mini, which replaces GPT-3.5, starting today.

GPT-5, mini, nano

$10, $2, $0.4 in output

GPT-5-mini vs. other model comparison

Features of the 🏆 GPT-5-mini

📊 Degree of detail of concept:.

More detailed and precise concept definition

Extract even the finest elements: **⿻ (Unicode symbol)**.

Compound understanding combining Japanese and symbols

🔗 Accuracy of the relationship:.

Six relationships are extracted (other models have about four)

High confidence (0.90-0.95)

Interrelationships are also captured (uses ↔ example_of).

https://gyazo.com/851cbee2aebdb5d10fc28abccbdef89a

I'd like to synthesize the two-way arrows because it's hard to see them overlapping.

It's a pre-LLM idea to extract a relationship like part_of in the first place, let the relationship be explained in words too.

---