Language is a medium that humans use for sharing thoughts, attitudes, and feelings. In order to do so it has to be embedded in some physical medium, whether sound, visual signs, of physical gestures. But the thoughts, attitudes, and feelings are not in the physical medium, they are not attached to the physical tokens in the way that the handle is attached to the head a hammer. Rather, they are associated with the physical medium through very complex and diffuse circuit patterns in people’s brains.
But the fact that we embody those thoughts, attitudes, and feelings in a physical medium allows us to get “meta” to them, to treat them as objects. How/why? Because the physical medium is an object that we can perceive like any other physical object. We can perceive the speech stream like we perceive bird song, thunder, a babbling brook, or the sounds of a horse galloping on the plain. And so it is with writing and signing.
This gives rise to what Roman Jakobson termed the metalingual function of language, language about language. Metalanguage is enormously useful in various ways. On the one hand, for example, it allows to “debug” conversations in process, thereby improving communication. But it also allows us to bootstrap abstract conversation over strings of dialog, as David Hays has argued (in e.g. On “Alienation”) And to evolve ever more sophisticated conceptual systems over the long course of history (Benzon and Hays, The Evolution of Cognition).
With that in mind, I want to take a look at ChatGPT’s grasp of metalinguality, if you will. First, I ask it about Jakobson’s idea. Then I probe its ability to “think” about language.
The metalingual function
What’s going on here seems self-evident to me and is not surprising. The web has many texts about Jakobson and his ideas, so ChatGPT had material to work from.
Do you know about Roman Jakobson, the linguist?
Yes, Roman Jakobson was a Czech-born Russian and American linguist and literary theorist. He was one of the founders of the Moscow Linguistic Circle and a key figure in the development of structuralism and the structural analysis of language. Jakobson's work on language and communication had a significant influence on fields such as linguistics, semiotics, and communication studies.
What did Roman Jakobson mean by the metalingual function of language?
In his work on the functions of language, Jakobson proposed that there are six different functions of language, each serving a different communicative purpose. One of these functions is the metalingual function, which refers to the use of language to talk about language itself. This can include things like discussing the rules of a language, defining words, or discussing the structure of sentences. The metalingual function is important because it allows people to communicate about the tools they are using to communicate, which in turn allows for a deeper understanding and more effective use of language.
Can you give some examples of language that illustrates the metalingual function?
Sure, here are a few examples of language that illustrates the metalingual function:
- "The word 'dog' is a noun."
- "In this sentence, the word 'big' is an adjective modifying the noun 'dog.'"
- "The past tense is formed by adding -ed to the base form of a regular verb."
- "A preposition is a word that is used to show the relationship between a noun or pronoun and other words in a sentence."
- "Do you understand the difference between a direct object and an indirect object?"
In each of these examples, the speaker is using language to talk about language itself, in order to clarify or explain some aspect of the structure or usage of the language. This is the metalingual function in action.
Self-referential sentences
These interactions are tricky, and therefore somewhat more interesting.
Count the number of the words in the following sentence: “John helped Jane to finish her homework.”
There are nine words in the sentence "John helped Jane to finish her homework."
[2nd try]
The sentence contains eight words.
[3rd try]
There are nine words in the sentence: "John helped Jane to finish her homework."
I learned a bit later that ChatGPT treats line-end punction as a separate word. Given that, it’s second try is correct, but the first and third are not. Why does it get the count wrong even allowing for the fact that it’s counting periods as words?
Continuing on with the session: On the off-chance that it was confused by the quotation marks, I tried again without them:
Count the number of the words in the following sentence: John helped Jane to finish her homework.
The correct sentence would be: “John helped Jane to finish her homework.” There are nine words in this sentence. It is important to include punctuation in a sentence, as it helps to indicate the structure and meaning of the words. Without punctuation, a sentence can be difficult to understand and may be ambiguous or unclear.
What it offers up as the “correct” sentence is identical to the one I gave it. It still counts the words incorrectly, and it seems to be asking for unnecessary internal punctuation.
Moving directly on (same session):
What are the parts of speech of the words in the following sentence: “John helped Jane to finish her homework.”
In the sentence "John helped Jane to finish her homework," the parts of speech of the words are as follows:
"John" is a proper noun, which is a type of noun that refers to a specific person, place, or thing.
"helped" is a verb, which is a word that describes an action or state of being.
"Jane" is a proper noun.
"to" is a preposition, which is a word that shows the relationship between a noun or pronoun and other words in a sentence.
"finish" is a verb.
"her" is a pronoun, which is a word that takes the place of a noun in a sentence.
"homework" is a common noun, which is a type of noun that refers to a general class of objects or people, rather than a specific person, place, or thing.Note that in this sentence, "to" and "her" are functioning as infinitive markers and possessive determiners, respectively, in addition to their roles as prepositions and pronouns.
How many words in this sentence?
I'm sorry, but I cannot answer that question as it is incomplete and does not provide the sentence or text that I am supposed to count the words in. Can you please provide the complete sentence or text that you would like me to count the words in?
That sentence is referring to itself. How many words does it have?
The sentence "That sentence is referring to itself" has seven words. These words are: "That," "sentence," "is," "referring," "to," "itself," and "." (Note that the period at the end of the sentence is also counted as a word.)
Is this sentence grammatically correct?
I'm sorry, but I cannot answer that question as it is incomplete and does not provide the sentence or text that I am supposed to evaluate for grammatical correctness. Can you please provide the complete sentence or text that you would like me to check for grammatical correctness?
Its response to the first prompt was fine. But it is not clear to me what’s going on from the second prompt on. There is some confusion. Consider its response to this prompt: “How many words in this sentence?” That sentence certainly is complete and its words countable, but perhaps ChatGPT didn’t get it. Perhaps it thought “this sentence” referred to some other sentence, which it was unable to locate [keep in mind that it keeps track of an entire session].
Now consider its response to: “That sentence is referring to itself. How many words does it have?” I was of course referring to the sentence in the previous prompt (“How many words in this sentence?”). It doesn’t realize and counts the words in the first part of the prompt, “That sentence is referring to itself.” Is that a reasonable thing to do? That’s certainly not what I had in mind, but: How would a human deal with those series of prompts? I can imagine that (some) real people would get confused as well. Then we have its response to the last prompt. Did it think the phrase “this sentence” referred to some other sentence which it couldn’t locate. Quite possibly.
I think this merits further investigation. But my head aches at the thought of the contortions that might involve. Do I prompt it with something like: “How many words are in this sentence right here?” or “How many words are in this sentence right here not some other sentence?” I don’t know.
Language is tricky. More later.
