Make automatic keywords more accurate
Make the keyword model more accurate: judge some sample photos yourself, and Portman2 suggests how sure it must be before a word counts, overall and word by word.
That is the cut-off. Lower finds more and is wrong more often. Open Refine the cut-off with Refine cut-off… under Machine learning; it opens in a window of its own.
Judging
The window shows one sample photo at a time, large, with the words the model suggested for it as chips beneath.
- Click a word that is right.
- Click it again if you cannot tell — it is left out of the counts.
- Click a third time to clear it.
A word you leave unmarked counts as wrong when you go on with Next image. Skip passes over a photo without judging it.
The suggestions
Press Done to see what your judgements add up to:
- Default cut-off — the default with the best balance of right words kept against wrong ones, compared with the default now;
- Words that leak — words wrong more than once above the cut-off, each with a suggested cut-off of its own, or removal;
- Never right, but harmless — words that were never right but stay below the cut-off.
Each suggestion can be ticked or not. Accept applies only what is ticked; Judge more goes back to judging. With only a few photos judged, the window says the numbers are not yet worth much.