[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [Elist Home]
Subject: Re: HM.Frameworks: Physical Description
I'm noticing on the discussion on arcs
(kindaLike, etc.) that we seem to be moving into a Fuzzy Logic (FL) realm. It
occurs to me that it may be worth either labeling or supporting arcs with an
intensity indicator as well [0,1] that would provide a mechanism for correlation
across a specific operator. For instance, if you have an arc operator of "Like"
then a like intensity of 0 would correspond to no correlation (the arc between
the two entities implies no correlation) while an intensity of 1 would imply
total correlation (1 is a synonym for the other) and an intensity of 0.7 would
indicate that there is a certain amount of correlation (the two are similar but
not the same). While for some of the arcs being dealt with these intensities are
largely subjective, the same can be said for HumanML in
general.
I'd push this approach for a couple of reasons.
First of all, it makes it possible to follow correlative links into a
hyper-linked environment, making it feasible to provide a point at which point
information ceases being reliably relevant. For instance, if you had an
arc:
Like at 0.7
A ---------------------> B
and another arc
Like at 0.8
B ---------------------> C
then you can view the arc A->C as being similar at 56% (which could be
interpreted as there is about a 1 in 2 chance that the two nodes are highly
similarly, or more accurately that there is a 56% correlation between A and C).
You could then follow a grove of nodes pruning those that fall below a certain
threshhold (say 0.1 correlation) as no longer being relevant to the current arc.
This also makes it possible to perform other analytical metrics (Bayes, for
instance) or to determine the mean relevance of a tree to a given node.
This is just an observation, do with it as you will.
-- Kurt |
[Date Prev] | [Thread Prev] | [Thread Next] | [Date Next] -- [Date Index] | [Thread Index] | [Elist Home]
Powered by eList eXpress LLC