5

The Context Principle says to "never ... ask for the meaning of a word in isolation, but only in the context of a proposition", which I take to mean that we should not look to words in isolation but to its use in the context of a true and complete proposition for meaning, whatever that actually means.

My question is how this is plausible, since we should obviously look elsewhere for meaning too. Someone just learning a language would never learn anything if you just gave them propositions in which words occurred; you must first understand those words and what they mean in isolation, and the grammatical structure of the language, for any sentence or proposition to be even understandable, let alone mean anything. Of course, you could restrict the principle and say that at least some words are understood through ostension or reference to experience, but then how do you even get anything like the Context Principle from this? You would be admitting that both experience/idea and propositions are important for many words.

In fact, what words could you understand without any such fundamental idea? Any definition or sentence like "A is B" or "A means B" would require you to say what B is, and until you refer to something outside propositions or sentences you never can convey meaning (since nothing would ever be understood to someone who does not already understand the language). So in what sense are propositions the actual carriers of meaning if basic cases of references are doing all the heavy lifting which enables you to understand anything in the first place? Of course, by using words in sentences we can better understand them, but every time I have ever come to understand a word this way, that contextual understanding is mediated through understanding other words, which would have to be mediated through others, etc. It is not that examples are not helpful, but they are only helpful when there is an idea or intuition there already.

My question is, then, how would someone defending the Context Principle respond? I feel as though there is some response here, but I cannot at all formulate it. Does anyone know of papers written about the Context Principle, particularly objections and responses to objections that are accessible?

Improve this question
9
  • The meaning of this sentence is not in the symbols, which are abstract, and have no inherent meaning: it is in the spaces between the letters and in the period at the end of this sentence. Commented Oct 22, 2024 at 1:09
  • 1
    @SystemTheory the word red, regardless of if you wish to use 赤 or красный to denote the experience, refers to something objective and meaningfully understandable. So the word does have inherent meaning outside of a sentence, and just because we choose some arbitrary symbolization does not imply the word is meaningless; I do not see a plausible reason for thinking so. Otherwise a blind person could understand the same meaning of the term red or 赤 or красный without ever seeing it, which seems odd. Commented Oct 22, 2024 at 2:20
  • 3
    @Curulian The word red refers to something objective and meaningfully understandable... communists! You can trip over 'em! Commented Oct 22, 2024 at 2:34
  • In the 2020 PhilPapers poll, 52% opted for moderate and 25% for radical semantic contextualism, so forms of CP enjoy broad support. I think you interpret Frege's CP radically, but what he is usually taken to mean is that only sentence's context determines the meaning of a word fully. This does not preclude loose commonalities across contexts that can be grasped piecemeal "in isolation" and composed to understand sentences, in agreement with semantic compositionality. Commented Oct 22, 2024 at 3:29
  • 1
    @MauroALLEGRANZA Babies learn to associate behavior with meaning. They try to mumble through speech until they hit a sound or two that are vaguely similar to the parents language that it individually gets reinforced as a word. They'll often use single words to express more complex thoughts, and you will see frustration when that single word doesn't elicit the wanted response. Babies definitely learn differently than they're taught. Commented Oct 22, 2024 at 15:05

3 Answers 3

Reset to default
9

Consider the following sentences containing the word 'apple':

  • I like to eat an apple from time to time. (a class of instances)
  • I bought the Apple computer. (a name of an organization used as an adjective)
  • You are the apple of my eye. (a metaphorical object)
  • 'Apple' is used to represent a type of fruit. (a token)
  • Apple threw me, as horses are wont to do. (a name of a horse)
  • Your apple is green and mine is red. (two physical pieces of fruit)

What does apple mean? In every example, it is has a distinct meaning, and the full meaning of the token 'apple' in every sentence is only understood within the context of the sentence. But, Conifold is right when he says that you are interpreting the Context Principle too stringently.

According to the correspondence theory of truth, a proposition is not just a string of characters; it is an expression that corresponds to a state of affairs. Thus, the maxim to look to the proposition isn't an admonishment to ignore the state of affairs in the world that the proposition represents, but to use the context of the sentence to determine which sense is applied to the token. From WP:

Gottlob Frege, Introduction to The Foundations of Arithmetic (1884/1980)... In the enquiry that follows, I have kept to three fundamental principles: always to separate sharply the psychological from the logical, the subjective from the objective... never to ask for the meaning of a word in isolation, but only in the context of a proposition... never to lose sight of the distinction between concept and object.

So, that is not to say that 'apple' doesn't have a meaning outside of sentence, but rather it has too many; the sense of the isolated term 'apple' is underdetermined because 'my apple' and 'your apple' and 'apple of my eye' are all apples that have different senses. In natural language processing, it's extremely difficult to interpret 'apple'. If one were to provide a series of comprehensions for the types of apples listed in the examples (class of instances, metaphorical object, etc.), it would be very difficult to automate detecting which sense of apple is applied.

The SEP's article on Compositionality (SEP) offers this, at least, as a way to understand the Principle. It says:

There is an alternative way to construe Frege’s principle, a way that makes it a determination claim, not a primacy claim. To state it in a form that matches the generality of (C) we should drop the talk of words and sentences, and talk instead about complex expressions and their constituents... (Fall) The meaning of an expression is determined by the meanings of all complex expressions in which it occurs as a constituent.

So semantic holism doesn't claim that words have no meaning, but rather, as the obverse of the Principle of Compositionality, meaning is conferred on smaller constituents by the larger structures they are contained in. This is one of the main points of Quine's argument about gavagai. Reference is not deductively certain, but rather is probabilistic, and therefore it is the context clues which actually lead the hearer to a likely, but uncertain meaning for each token in the process of interpretation. (And to boot, paragraphs affect the meaning of propositions, and so on...)

Improve this answer
3
  • BTW, if you're interested in the Context Principle, a good place to start above and beyond the SEP articles on Frege is Dummett's work: Frege: Philosophy of Language. Commented Oct 22, 2024 at 14:52
  • General intelligence is the innate or inherent ability to comprehend human or human-like meaning. Such intelligence is distributed among human subjects. Language is an effort to encode and decode subjective meaning with the automatic assumption that other humans are creatures like oneself. Commented Oct 22, 2024 at 15:31
  • @SystemTheory That's certainly moving in the right direction for defining general intelligence. AI has been moving in that direction for 75 years. The first researchers were optimistic they could crack it in a decade, and now if AGI studies are any measure, it will be a long time before systems are sophisticated enough to "comprehend human or human-like meaning". Let's be honest. Many actual human beings struggle with comprehending certainly pockets of meaning, to be sure. Commented Oct 22, 2024 at 17:32
4

The context principle you mentioned can have distinct manifestations at various levels. On the most basic level, (1) you can have a term that has distinct meanings, and it is only from context that you can determine which meaning is intended. Alternatively, (2) you can have a situation common in technical fields like math or philosophy that a term is used in a technical sense different from its generic sense though often related to it to a certain extent. Most radically, (3) you can adopt a marxist perspective that a given term may have distinct meanings depending on whether it benefits the class of the oppressed or the class of the oppressors.

I recently came across a striking instance of situation (2) above, where Abraham Robinson, who is a Formalist, described infinite totalities as lacking a reference in either the physical or any Platonic realm. Sometimes Robinson used the adjective meaningless in a technical sense (different from its generic sense) to describe infinite totalities, always clarifying that what he meant was an absence of reference in the sense just explained (thus "meaning" is being used as synonymous with "reference"). On the other hand, a recent article analyzing Robinson's views:

Erhardt, R. Denying Infinity: Pragmatism in Abraham Robinson’s Philosophy of Mathematics. History and Philosophy of Logic (2024). https://doi.org/10.1080/01445340.2024.2344346

has mistaken the technical sense for the generic sense, resulting in a rather dramatic misreading of Robinson's philosophical position.

Improve this answer
3

The distributional hypothesis in computational linguistics (distributional semantics) is a better way to formulate the context principle. The hypothesis is simply that words that occur in similar contexts have similar meanings. Context here refers to the words preceding and following a given a word as used in actual texts. We don't need to stipulate how big that context is or if it corresponds to some preconceived linguistic (or other) notion (like "proposition").

This is better in the sense, that it doesn't need strict a priori definitions of what a "word", "meaning", "proposition" or "context" is and how those might relate to each other. It's also better in the sense, that this leads to empirical hypotheses (about people's sense of synonymy or meanings) that can be tested by constructing language models. As such, I don't think that anyone (at least nobody in computational linguistics) has any doubts about the fruitfulness of the hypothesis or about its general truth.

Based on this hypothesis we can contruct word embeddings where each "word" (if we use those as units) is represented as a vector of reals (floating point numbers) which encodes the various aspects of "meaning". It turns out that people's intuitive/acquired sense of synonymy (of similarity between "meanings") can to a very large extent be modelled extremely well with the formally defined similarity between those vectors (for instance, as cosine similarity).

This general hypothesis is also at least compatible with language acquisition by human children. Language is embedded in wider non-linguistic contexts of behavior and of purposeful action and problem solving. In the most primitive language games we use language to make others obey our commands or to entice them to do something for us or to not do something. (CIA courses for lesser known languages usually start with commands like "Stop!", "Show your hands!".) Something similar happens in the training of generative LLMs: the input embedding and the models are trained with respect to certain given tasks, so with respect to certain expected more or less correct actions (for instance, "fill in the blanks" or "give a summary").

Philosophers sometimes ask whether an LLM "really" understands or really "means" anything -- since the system doesn't have any purpose or intentionality. But they seem to forget that the intentionality is largely there from the very start: by training a system to respond as expected, so as intended by us, it may (and in fact does) start to respond as intentended. If so, then what more could one ask of it?

Improve this answer
2
  • +1 'Based on this hypothesis we can contruct word embeddings where each "word" (if we use those as units) is represented as a vector of reals (floating point numbers) which encodes the various aspects of "meaning".' Someone knows that one can renorm data by using the unit circle to optimize ML models. Every token has the potential to be monosemantic! Commented Oct 22, 2024 at 18:58
  • Many maps of meaning must exist in the human mind, as so-called general intelligence, to describe language models which relate words to vectors of discrete floating point numbers. Otherwise the language models, the words and vectors, are just meaningless data. Commented Oct 22, 2024 at 20:28

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.