I am wondering if there is a theory/field/framework in linguistics today which attempts to explicitly describe completely the intrinsic content of a “concept”.

Of course, concepts reside in the mind-brain so it would either need to be an “average” based on data, linguistic and also sensory (video, audio, etc, ideally all components of human experience, including registering bodily sensations, but that’s not currently realistic), of the use of concepts across multiple people (or all people), or it could be a study of a concept in the mind-brain of one person.

The more important thing is what we currently believe any concept to truly be constituted of.

I generally think of concepts as “clusters” - if we consider every “element” of being a conscious human:

  • sensory fields including vision, audio and smell with “algorithms” for organizing and “chunking” those inputs, i.e. not just a messy flux of senses but the ability to discern and separate regions of space into objects, sounds into having different sources, etc

  • knowledge, as well as the ability to think and reason, i.e. we create strict criteria for certain concepts

Basically we can imagine a concept in our mind is a sort of clustering of emotions, memories, experiences, images, and semantic knowledge or associations acquired for the concept. We might just have a general association between the word “democracy” and “freedom”, guiding the intuitive use or understanding of the word, but we might strictly know a country where everybody is imprisoned by a tyrant is not a democracy, or even more strictly it’s not a table if it isn’t an elevated surface you can place things on.

I guess I should focus more on that question of strictly semantic knowledge. We have to consider that semantic knowledge has to occur through bootstrapping; after learning your first words you pretty quickly transition into semantic knowledge; being able to explain that a hippopotamus is an animal that lives in Africa, etc.

I’m just trying to figure out how you could exhaustively list someone’s semantic knowledge on a concept. It might be intrinsically impossible because you could ask an original question which would require them to think and come to a decision, like “does hippopotamus meat taste good?”

Maybe you could try to delineate a boundary by specifying what the person considers knowledge they possess, i.e. facts. I’m worried that still wouldn’t work because people can be be aware that a concept is not perfectly defined yet still implicitly show an intuition of what it’s like.

I guess I’m just trying to think of a way to map out what a person’s “decision-tree” actually is with accuracy when faced with a picture or a description and asked, “Is this an X?”

The problem is we can imagine many criteria all existing in parallel. We could claim that the more of the criteria fulfilled the higher the likelihood. So if it has the color of a hippopotamus, the size of a hippopotamus, each body part the shape of a hippopotamus, if it’s somewhere a hippopotamus would be, all of these things make it more likely to fit the prototype. If a few of them are lacking it wouldn’t matter. If many are it would start to fail to resemble the prototype enough.

The problem is that some of the criteria are not optional, they are really strictly determining, like a fake gun is not a type of gun, even if it looks exactly like one.

So I can’t imagine how to diagram that. Some of the factors could be given weights and values. I.e., as to the question of if you are overweight, the value would vary continuously (a number), whereas the weight of the value would be very high: it is a strongly determining factor in the decision.

Maybe the binary factors could have Boolean values.

Maybe you could structure it like all known determining factors come first in the decision tree, followed by more variable factors with probabilities?

I cannot see the physical form of such a diagram but I am actually thinking of a neural network possibly being able to support that kind of decision process. But the problem is that you can’t explicitly see what the nodes in the network may correspond to conceptually, i.e. if the network has figured out that oranges are always a certain color, where in the network is the node that is representing that decision?

So, how can you unify logical criteria with prototype “likelihood” attributes?

And can we use explainable AI to observe the inner functioning / semantic knowledge of a large language model, to see how it might form myriad, complex decision graphs based on simultaneous criteria, to see what those criteria are?

1 Answer 1


This is mostly a philosophical question ("what is knowledge"), not a linguistic one, but it's an issue that may be in part addressed by at least some linguists. The question as posed presupposes a particular kind of answer, where I think most linguists would not whole-heartedly accept those assumptions, so instead I think the question should first be "What is a concept" – raise the question of ontological status first. If we can settle on what a concept is, existentially, then we might also be able to raise questions about how we learn them and how we use them. So this is the question of what a concept is "constituted of". A sensory impression is not necessarily part of a concept, it may well just be something that plays a role in creating a concept.

This article lays out the main theories of what concepts are, at the hands of philosophers. One view is that they are mental entities, understood in a representational theory of the mind. A second view is that a concept is an ability of a cognitive agent. The third view is that they are Platonic abstractions. I believe that the majority of linguists who think about this matter assume the representational theory, although I also think that linguists tend to think of concepts as being representational objects that can be used to do certain things. And furthermore, most linguists have a mildly Platonic streak that allows them to set aside unanswerable questions about brain states and quantum mechanics. There is probably an increasing tendency for linguistics-adjacent experimenters to devise approximations of actual neural nets (if such things exist) via computer programs called "neural nets" (i.e. artificial neural networks).

There are two main approaches to concepts: the crisp, and the sloppy. The crisp approach says that there are necessary and sufficient criteria for categorizing things, so that you can say if a thing is a "snake" or a "worm": concept have definitions. The sloppy approach says that there is a probability function between observations of an individual and a classification, thus a pigeon is maximally like "a bird" and a penguin is least like "a bird". The continuum of uncertainty can be such that we really can't say if a dingo is a dog, or is it dog-adjacent, and the discussion devolves into the truism that "it depends on how you define X".

Linguists tend to favor the sloppy approach, meaning that there are many marginal cases but vastly more clear cases. In the crisp view, the observed behavioral data is explained in part by the fact that all representational objects in the mind (like a definition) are in the mind of an individual, and individuals can differ. Plus, individuals are generally aware that other individuals exist and may hold different opinions, therefore their behavior may reflect knowledge of social variation. For example, I am aware that there are different theories of "concept", so my behavior does not necessarily rigorously reflect my own ideas about what a concept is.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.