After running this on many different pieces of text, I got no score lower than 0.7106113. So naturally it would make sense to assume that the output is between 0.7 and 1. If the score were calculated by Math.max((vector.score - 0.7) / 0.3, 0), the score would be a number between 0 and 1, and a threshold of 0.85 using the original formula would be 0.5 (50%) using this adaptation, which makes more sense. The score is also clamped at 0, so if somehow the score was less than 0.7, the score outputted would never be negative.
If you wanted 50% to be equivalent to the current 86% threshold, you would do Math.max((vector.score - 0.72) / 0.28, 0).
It may be useful to return this in a normalizedScore / normalisedScore from the API so as to not break anything currently using the score property.
After running this on many different pieces of text, I got no score lower than 0.7106113. So naturally it would make sense to assume that the output is between 0.7 and 1. If the score were calculated by
Math.max((vector.score - 0.7) / 0.3, 0), the score would be a number between 0 and 1, and a threshold of 0.85 using the original formula would be 0.5 (50%) using this adaptation, which makes more sense. The score is also clamped at 0, so if somehow the score was less than 0.7, the score outputted would never be negative. If you wanted 50% to be equivalent to the current 86% threshold, you would doMath.max((vector.score - 0.72) / 0.28, 0).It may be useful to return this in a
normalizedScore/normalisedScorefrom the API so as to not break anything currently using thescoreproperty.