Natural Language Understanding and the Meaning of ChatGPT - Einzelansicht

Please note: The first meeting is on April 25th, i.e. in week 2 of the semester.

How does a sign on a billboard use language? It doesn't; whoever put writing on the sign is using the sign to use language. How does your computer use language when it shows you an error message? Again, it doesn't; the programmer who connected the string to a triggering condition does. How does a parrot use language? Again, arguably, it doesn't, although it gets harder to explain why, and what is even happening when a parrot vocalises something. Finally -- how does a Large Language Model use language?

This course will deal with the question of how best to characterise what these new computer systems do with language.
To make matters concrete, we will look at situations like the following. Suppose you typed into a cLLM ("chat-optimized LLM", such as ChatGPT) the string
"What happened in November 1963 in Dallas?", and you received the output
"In November 1963, President John F. Kennedy was assassinated in Dallas, Texas."
What is the connection between this production of that string, or of that sentence, and the state of affairs of JFK having been assassinated? What needs to be the case for such a production to be justifiably understood as an assertion, and for the name "John F. Kennedy" to refer to the historical person (with whom very few speakers today will have been personally acquainted)?
What does it mean, if anything, if a system that has produced this string (which we would take to be a true claim) is also liable to produce strings that evaluate as false, or if anything changes if it becomes less liable to do so.

Thinking about these question will take us on a long journey, starting from the (quite confused, as we will see) public discussion about such systems; over the increasingly helpless attempts at measuring what the programs do; to a deep dive into theories from linguistics pragmatics and the philosophy of language about what *we* do when we use language; and back to attempts at understanding LLMs---and ways of using them without being used by them.

This is not a course about building Natural Language Understanding systems. We have other courses where that is the topic. Nevertheless, the course requires a good technical understanding of LLMs---as well as the ability to read and critically engage with literature from a variety of fields.