Office Hours Getting to know... Simon Charlow

Linguist Simon Charlow’s journey to Yale was a circuitous one that included a detour into private industry.

After spending about eight years on the faculty at Rutgers University, the formal semanticist and computational linguist took a leave to work for a Boston-based semiconductor company.

“I wanted to experience how linguistics translated from academia to industry,” said Charlow, who this fall joined Yale’s Faculty of Arts and Sciences as an associate professor of linguistics. “The first thing I learned was that my new colleagues were phenomenally smart and interested in the same kinds of questions that fascinate me.” 

At Yale, Charlow is applying aspects of mathematics and computer science to better understand the processes through which human beings encode their thoughts and ideas into language.

In the latest edition of Office Hours, a Q&A series that introduces new Yale faculty members to the broader community, Charlow discusses the interplay between linguistics and computer science, the influence of AI on his field, and what attracted him to New Haven.

How do you describe your work to people unfamiliar with linguistics?

Simon Charlow: Most people don’t know what linguists do — it’s not like it’s taught in high school. We study language through a scientific lens. We’re interested in the sound patterns of languages, how people learn languages, and how languages become encoded in your brain. We think about the syntax of languages, which includes the grammatical rules governing the order of words in sentences.

I’m a semanticist, which means that I study meaning in language and how complex meanings are built out of simpler pieces. Unlike basically every other creature on Earth, humans can take something that’s very private — a thought or realization — encode it into language, and transmit it into other people’s heads. I use tools from math, logic, and computer science to study how that encoding and decoding process works.

Where do linguistics and computer science intersect?

Charlow: I’m not a computer scientist by training, but I became interested in computer science as a way of shedding light on fundamental questions about language and linguistic meaning. In a very basic sense, computer scientists are interested in certain kinds of language, not spoken languages that evolve naturally, but programming languages created by humans. There are programming-language theorists who are interested in the syntax and semantics of computer languages, which may appear very distant from how natural languages work. 

In my work, I’m very interested in identifying commonalities, sometimes latent or non-obvious ones, between these artificial languages we’ve invented to program computers and natural-language meanings, which nobody invented, per se. What can we learn about how particular meanings are constructed in natural languages by studying how programming languages encode meanings?

Have recent developments in artificial intelligence affected the study of linguistics?

Charlow: There is an interesting tension these days between the symbolic models that linguists have used to help us understand so much more about language than we did 50 to 70 years ago and the large language models [LLM] that drive ChatGPT and other generative AI tools. These symbolic models aren’t really in the same ballpark as ChatGPT in terms of their ability to process language. What should we make of that? On one side of the conversation, people argue that LLMs are so good that we don’t really need linguistics anymore. I obviously disagree with that, and I believe that symbolic methods remain an important part of the story of how language works.

For one thing, LLMs don’t learn very efficiently. When children learn languages, they do it with very little data compared to LLMs. An LLM requires several orders of magnitude more language — at least 1,000 times more — than what a young child hears from their infancy to age 5, so humans are far more data efficient. Also, there are aspects of meaning and language use that LLMs struggle to capture, like reasoning deductively to solve problems in domains outside what they have been explicitly trained on.

But I think it’s a two-way street. Not every linguist would agree with me here, but I think LLMs can be used to inform the study of human language. Maybe they embody important insights about how people learn language, or about how certain kinds of linguistic meaning are represented. Maybe they can shed some light on how language evolved.

What attracted you to Yale?

Charlow: I took a circuitous route to get here. I started at Rutgers, which has a great linguistics department, and then I switched to private industry; I worked for a Boston-based semiconductor company for about two years prior to coming here.

Yale’s linguistics department has an incredibly strong faculty that is producing empirically informed, theoretically sound work on a range of important questions. One thing that sets Yale apart from other institutions is its strength in theoretically informed computational linguistics. Bob Frank [professor of linguistics] and Tom McCoy [assistant professor of linguistics] are doing fascinating work in this area. Both are pursuing some of the questions involving AI that we touched on earlier.

Another big draw was Yale’s amazing philosophy department where faculty are doing important work on the philosophy of language.