There should be a way to designate negative vs. positive, which doesn’t rely on word similarity, since their meanings are, though on the same dimension, oppositional to one another.
What kind of a mathematical function would represent this general idea? For instance, the positive would likely be addition/multiplication. You’re moving right on the number line. The opposite, in relation to the positive, would move left on that number line. And in so doing the number line also becomes an axis — a vector, really, in the physics sense: with length and direction. So it certainly could be 4x + 6. Or it could be 4x or 3x. Which? Can a neural net figure out which? Can we present a bunch of whiches?
The RGB color code for orange is RED=255, GREEN=165, BLUE=0. Each pixel on a computer screen is composed of three small dots of compounds called phosphors surrounded by a black mask. The phosphors emit light when struck by the electron beams produced by the electron guns at the rear of the tube. The three separate phosphors produce red, green, and blue light, respectively.
It should be possible to apply physical combinatorials to similar concepts. For instance, to say something is ‘a blue pixel’ just means it’s atomized. r,g,b. It’s b, and that means that r and g are 0, and b is 255. If you say an orange pixel, it really means RED=255, GREEN=165, BLUE=0. The “atoms” (the indivisible particles) may be represented as one-hot encodings as to whether or not they’re present. In the former case it’s 001 and in the latter it’s 110. But there’s more information there in the actual strengths of the light hues. Each of those can be considered an atom at the same time, though, but at a slightly higher level: you might need two pixels to make this new ‘atom’, aka orange. It is perceived by a human to be orange, on a continuum that has no fixed point at which, say, green mixed with red turns to orange. Rather there is a range, probably the normal distribution, between which most people would agree that it’s orange.
And so it is with concepts. This is the rule of generality. A concept is necessarily composed. Are there “base level” or “root level” concepts? How do you measure the complexity of a concept? Is the Hunger Games or Game of Thrones a concept? Not exactly. It is a composed thing, but it’s a narrative. If you want to teach a computer both narratives and concepts, how do you do it? You could potentially use brighttext for that. But you must have changepoints within changepoints, I think. Would it help to input A Harry Potter book but with alternatives swapped in such that you could say, “This is bad, this is good so you show us the good?” That is by tagging the incorrect alternatives; by saying that Rowling’s combinations are better than anyone else’s? You’re essentially teaching it what’s good and what’s shit. That’s a taste, an opinion, a subjective experience.
Consider the following stringclump: amazed, bold, porcupine, Saskatchewan, whatever. “Whatever” is a lot more general than the others. Here, it validates the listener’s sense that these words have nothing to do with one another. Specificity is a dimension. 1956 Aston Martin is very different from car, or antique car, in a way that it isn’t from 2010 Prius. Although that’s also different. The point is that even differences are differentiable from one another. Some are along the same axis, and some are not. Again we come to this idea of a vector whose path diverges, or simply extends, or even goes into another dimension. The boy expired after a week in the hot room. Boy is young, but the heat in the room is another dimension which caused him to expire.