Source: Watercolor painting - Wikipedia

Caveat: the maps below are likely wrong. Please don't use them as examples of good concept maps. They're only shared publicly to document the tool's evolution.

First Try

Watercolor painting - A concept map by LLMapper

Mostly accurate, but suffers from having multiple root concepts. Would be ideal if the first pass could identify a central concept.

Sixth Try

Watercolor painting - A concept map by LLMapper

Much better. Some issues, which I suspect are rooted in problems at the summarization step. Inaccuracies. (Paper is made of watercolor?)

Fourteenth Try

Watercolor painting - A concept map by LLMapper

Structurally better, even though there are inaccurate relationships. (Paper handles handle?)

Comments

Today’s experiments were a failure. I suspect there’s something particular about certain kinds of Wikipedia articles that lead to multi-root diagrams. The only way I’ll be able to debug this is by exporting the output of the first prompt.

Changes

  • Instructed first prompt to have the object of every sentence except the first consist of the object or subject of at least one other sentence in the set.
  • Instructed first prompt on subject-predicate-object structures.
  • Instructed first prompt to conflate some concepts to minimize their number.
  • Modified LLMapper to export the output of the summary step and prompt 1.
  • Modified the summary prompt to make it more explicit.
  • Switched model for the summarizer — this degraded output, so I switched back to GPT-4.
  • Standardized around U.S. spelling.
  • Standardized concepts on lowercase (except when they are proper names.)
  • Instructed second prompt to only use the sentences in the source text.
  • Removed instruction to keep concepts to a minimum.
  • Instructed second prompt to remove pairs that include verbs as concepts. (May regret this change down the line.)
  • Increased sleep time, since I was getting lots of errors.
  • Removed instruction from second prompt that caused GPT-4 to behave like a software developer.
  • Explicitly asked the first prompt to pick an existing subject or object.
  • Instructed second prompt to always enclose concept names and labels in double quotes.