- From: Charles McCathieNevile <charles@sidar.org>
- Date: Tue, 19 Apr 2005 20:15:06 +0200
- To: "Nils Ulltveit-Moe" <nils@u-moe.no>
- Cc: "Giorgio Brajnik" <giorgio@dimi.uniud.it>, public-wai-ert@w3.org
On Mon, 18 Apr 2005 21:07:31 +0200, Nils Ulltveit-Moe <nils@u-moe.no> wrote: > Hi Charles, > > I would suggest that confidence is interpreted as it is defined in > statistics; i.e. as a probability which is a real number from and > inclusive 0 and to and inclusive 1. ... > Using the probability is a generally accepted way of representing > confidence values in most areas of science. Other representations can be > derived from this. Well, if the probability can be statistically determined (which happens often enough for it to be considered a useful convention) this makes sense. But, for example, the kinds of confidence that people are talking about when looking at alt text values are really pretty random - certainly not accurate to 1 whole significant figure. These could be refined over time by applying machine-learning principles to a large setof human-monitored results, giving something statistically more useful. But starting from where we are now, I think that we need some wayto say that one setof confidence values doesn't necessarily map cleanly to another one, and it seems to me that datatypes are a good way to do that. It might be sensible to define a datatype that is a probability value, and tell people that unless they are applying a known algorithm with known means of determining confidence they should subClass it for the datatype of their confidence. Which brings us back to the issue that we have to wait a little longer to get the best ideas of the best RDF developers on the best way to define datatypes... mvh Chaals -- Charles McCathieNevile Fundacion Sidar charles@sidar.org +61 409 134 136 http://www.sidar.org
Received on Tuesday, 19 April 2005 18:15:22 UTC