# What is an objective criterion for “specificity”?

I was trying to ask ChatGPT to be more “specific” and it made me wonder what an objective criterion for “specificity” is, given that I found it slightly hard to formulate.

All I can say is that concepts seem to exist in a hierarchy of subcases and superclasses - hyper- and hyponyms. And yet… who’s to say there’s a “bottom”, a “most specific”? I think the issue I faced is that specificity is easier to understand or define only relatively. I think in mathematics an “affine space” is a space with a concept of distance and order but no fixed “origin” point - you have more and less, but there is no central point of reference, no beginning.

I find it philosophically intriguing because one might say that you can always break a concept into a few if you just think about whatever properties or attributes it has and subdivide accordingly.

But this still seems like it would bottom out once you run out of concepts, in your language/worldview/conceptual framework/understanding of something. You could just keep taking the intersection of more and more concepts you already have, like, “a monkey who reads quantum physics books written by an orthopedic surgeon married to a Brazilian kite surfer who is blind and can read from right to left and upside down because they practiced reading in a mirror when they were younger…” and so on.

But that is not the only way to increase the specificity of a concept - color is a good example where by seeing changes in a phenomenon itself we can identify those… states, conditions, qualities, and give distinct names to them. So while concept “appending” is valid yet slightly trivial, the creation of raw concepts from the phenomenal world itself is a bit more tricky. I think it basically requires the observation of a patterned phenomenon; and the limit to which it can be made more specific is infinite so long as that phenomenon itself is infinite in scope; or put better, the amount of specificity possible is a direct mapping of the inherent form of that thing. One can observe light in smaller and smaller levels of precision, giving a “name” to each difference in wavelength (be it the number of that wavelength, or just an original name) - yet it will end as one arrives at the energy quantum (?) - the smallest unit of energy physically possible.

Then maybe you can provide “objective” instructions for someone (or a machine, or an algorithm) to make something more specific by asking it to name all possible subtypes, and then to select the subtypes it thinks are relevant to some situation at hand, being discussed. Still, it might be difficult for a machine that only knows words it was trained on, and may or may not be able to conceive, logically, or, requiring perception, that “red” can be subdivided into finer and finer gradations, just by perceiving them.

It seems “specificity” could be modeled/defined linguistically as the “size” of the set of such elements, possibly - the smaller, the more specific. There are less clowns who like baseball than clowns, unless every clown likes baseball. Or rather - because it is not about real-world incidence of said thing - there are less known things (in a predefined system of concepts) which would qualify as included in the set “clowns who like baseball” - the set would include “clowns who like baseball and are vegetarian”, but not strictly “clowns”. So then: if the set of (elemental, primary, non-composite) concepts we have to begin with is a specific number (m), we can combine the primitives a very large number of times (m choose n).

I still think this is not a complete answer because there are paradoxes appearing in my mind, for example, “red” is more specific than “thing”, but “red” would appear to be more of an elemental concept than “thing”. Maybe we have to start with the conceptual primitives and somehow we allow ourselves to generate composite concepts in both directions: generalization / abstraction by collecting sets of primitives we think are alike (“red”, “yellow”, “green”), and enshrining that set with a name (“colors”) - and maybe increasing specificity with the intersection of concepts. Maybe we could objectively define specificity like prime decomposition: the more “prime concepts” a composite concept has… the more information it has determining it - and the more “specific” it is?

I leave this as an open question.

• Machine learning has the concepts of "precision" and "recall" (en.wikipedia.org/wiki/Precision_and_recall). But it seems you have questions about knowledge modelling too, which is an old topic in artificial intelligence - no idea where it's at today. Jan 27, 2023 at 20:12
• Jan 27, 2023 at 22:45