https://arxiv.org/abs/2208.02957
Found a very new and interesting article about LLMs that could be read as a direct response to Barry Smith’s articles and even to his new, upcoming book (2022). For me, a philosopher, this article is particularly interesting because the authors are trying to make a philosophical case about meaning or even about language in general.
Contrary to what many people have claimed, namely that LLMs “will never achieve ‘meaning’ or ‘understanding’,” they suggest (1a) that LLMs can gain something like “meaning” and (1b) that LLMs already operate with human-like meaning in “many (but not all) ways” (p. 1). I was curious to read what the authors understand as “meaning” in this case when they claim that LLMs “achieved some key aspects of meaning” (ibid.).
The article begins with a reference to Bender and Koller (2020) which seems to me, to my point of view, very intuitive. Bender and Koller claim that machines do not mean what they say because they lack the reference to objects. Machine learning (training) on text alone is not sufficient to acquire meaning, because meaning is linked in some shape or form to objects in the “real world” (think of Wittgenstein’s many examples of how language-learning works). Bender and Kollwer call this the “octopus test”: “The octopus test assumes that reference determines meaning” (p. 1.). I do not want to summarize the very illustrative octopus test, which I like a lot. What is more interesting for me is the argument (ad authorities(?)) that follows as a response from the authors of this article to the octopus test:
“The octopus test assumes that reference determines meaning, but in fact cognitive scientists and philosophers have found a variety of problems with this view. One is that there are many terms that are meaningful to us but have no discernible referent at all, such as abstract words like ‘justice’ and ‘wit.'” (p. 1-2)
This response is funny to me because it is simply an instantiation of this never-ending claim coming from an ontology of objects: as if terms need an object that you can touch for them to be meaningful; and if you can’t touch “justice,” then this term, because it has no object, has no meaning.
Well, if the world consists of more than objects than we can touch, feel or see, than we’re going right back to the octopus test, claiming that terms, in order to have meaning, need reference. There can be reference, for example, to a socially shared and established practice or to something that is executed at the court, through institutions, etc. etc. The term “justice” has meaning for us, in our use of the term, because we refer to something that is real in our world. The octopus, lacking this reference, could maybe learn to use the term “justice,” but he wouldn’t know the meaning of the term.
The authors then try to show that reference can be “decoupled” from meaning. Their examples are: a treaty or a contract: “a treaty is still valid if the piece of paper is destroyed” (p. 2). I am actually not sure what this second response/argument is meant to demonstrate. That there can be meaningful references even though there is no actual “concrete referent” (p. 2)? But what does this have to do with the destruction of the piece of paper? The reference to Frege’s famous example of “morning star” / “evening star” doesn’t help me here because it still just reiterates the presumptions from the “ontology of objects”; namely, that terms can only have meaning if they refer to existing objects (objects I can see, touch, test, verify). What the octopus test shows, and what the authors fail to see (in my opinion), is that meaning is based on a practice (Wittgenstein’s famous “Gebrauchstheorie der Bedeutung”) – on how the term is used in a society. If we use a term to refer to XYZ then the meaning of the term is this reference. XYZ does not have to exist. Reference and meaning are still intertwined.
The next argument is quite funny too. It is, in fact, just another strawman argument in the shape of: “Iff we assume that the world exists only of objects,…” and “iff the term ‘postage stamp’ does not refer to an actual object”, “then meaning has nothing to do with reference.” Let’s see what I mean by this:
The authors describe in some detail how we use terms like “postage stamp” and they notice that we can understand the meaning of the term without having “considered all of the possible referents […], so reference cannot be what determines the concept” (p. 2). I do not even believe that the “octopus test” was meant to show that reference determines the concept; more like, meaning has something to do with reference. But that is not even the main point here. The authors seem to get to an understanding that humans do not need a list of all the possible referents of a term in order to understand the meaning of a term. They just give an example of how meaning works. Meaning is not a fixed set of definitions, but something that can change over time – according to how terms are used in the society. In my opinion, with this second argument the authors just showed (by accident?) that meaning has to do with reference and that this reference is based on our social practice. Another point for Bender and Koller (2020) and for the octopus.
The authors then even refer to Wittgenstein’s Philosophical Investigations (!), apparently without having read Wittgenstein. They use Wittgenstein for their argument that reference “plays little or no role in determining meaning”. Wittgenstein only says that the referent plays no role in it – the reference is still essential…
The authors explain their observations about meaning in the following sentences, saying that the meaning of concepts is “intertwined with other concepts.” But all these sentences are non sequiturs and therefore highly problematic. I have to quote it in length:
“One way to explain these observations is to assume that our meaning for terms like “postage stamp” (or “water”) may be primarily determined by the role these concepts play in some greater mental theory. Very roughly, people call something a “postage stamp” if you pay for it and then attach it to a letter in order to have the letter to be delivered. In this view, the meaning of the word is intrinsically intertwined with other concepts like “payment”, “letter” and “delivery.” (p. 2)
Notice the apparent jump from “call something a ‘postage stamp’ if you pay for it and then attach to it…,” which is, if I am not mistaken, a description of a social practice and of how people do something, to “in this view, the meaning of the word is intrinsically intertwined with other concepts” (ibid.) which suddenly seems to be purely linguistical?
The Wittgensteinian point here is that how we use terms determines the meaning of a term. Obviously, how we use terms is intrinsically intertwined with other concepts, but this “intertwinage” (what is the noun of being intertwined?) is not the constitutive element of the meaning – it is only a consequence. And this link between concepts can only work because, primarily, we use words in a certain way: we use words to refer to “things” (all things that you can imagine).
The authors then conclude what they cannot and shouldn’t conclude:
“Such relationships between concepts are the essential, defining, aspects of meaning and, in fact, possessing the appropriate relationships allows you to determine the reference. This view makes sense
of the puzzling examples above.” (ibid.)
No, no, no. That is not at all how you can make sense of the examples. That’s not even how you yourselves described and explained the examples. The authors themselves were referring to practices that these words are referring to. The meaning of the word “postage stamp” is not its interrelation with other words like “payment” (that’s what the authors claim) but its interrelation with things; with other words, its reference. We understand the meaning of “postage stamp,” because we know (roughly) how people normally use that word. Only on the basis of this social practice, there can be an interrelation of concepts. We know the meaning of “postage stamp” because we know what people are referring to when using that word. The octopus who only learns how to use these words in sentences “I have to pay for the postage stamp. I have to attach the postage stamp to a letter. I need to attach it before sending it away.” will never know what these words mean even if it “knows” the (most probable) interrelation of “postage stamp” with other concepts like “payment,” “letter,” and “delivering.” (ibid.)
I have to say that this whole paragraph (and the reference to Wittgenstein) is really, unfortunately, blatantly wrong. And Wittgenstein’s Philosophical Investigations is, in fact, the best case against their semantic, linguistic, representational theory of meaning. They call this approach: “Conceptual role theory.”
I continue reading tomorrow, if I find the time. (August 09th, 2022).
When do you plan to continue this?