For the first time, AI analyzes language as well as a human expert
The original version from This story appeared in Quanta Magazine.
Among the countless abilities that humans have, which ones are unique to humans? Language has been a prime candidate since at least Aristotle wrote that humanity is “an animal that has language.” Even as large language models like ChatGPT superficially replicate normal speech, researchers want to know if there are certain aspects of human language that simply don’t exist in the communication systems of other animals or artificially intelligent devices.
In particular, researchers are investigating the extent to which language models can reason about language itself. For some in the linguistic community, language models are not alone don’t They have the ability to reason can’t. This view was summarized by eminent linguist Noam Chomsky and two co-authors in 2023, when they The New York Times “The correct description of language is complex and cannot be learned just by putting it in big data.” These researchers argued that AI models may be adept at using language, but they are not capable of analyzing language in a sophisticated way.
This view was challenged in a recent article by University of California, Berkeley linguist Gasper Bagosh. Maximilian Dabkowski, who recently received his PhD in Linguistics from Berkeley. and Ryan Rhodes of Rutgers University. The researchers put a number of large language models, or LLMs, through a range of linguistic tests—including, in one case, the LLM’s generalization of the rules of a fictitious language. While most LLMs were unable to parse the language rules in a way that humans could parse, one had remarkable abilities that far exceeded expectations. It was able to analyze language in much the same way that a graduate student in linguistics would—drawing sentences, resolving multiple ambiguous meanings, and using complex linguistic features such as recursion. “This finding challenges our understanding of what artificial intelligence can do,” Beguš said.
The new work is both timely and “very important,” said Tom McCoy, a computational linguist at Yale University who was not involved with the research. “As society becomes more dependent on this technology, understanding where it can succeed and where it can fail is increasingly important.” He added, linguistic analysis is an ideal test bed to evaluate the degree to which these language models can reason like humans.
Infinite complexity
One of the challenges of giving language models a rigorous linguistic test is making sure they don’t already know the answers. These systems are typically trained on vast amounts of written information—not just the bulk of the Internet, in dozens, if not hundreds, of languages, but things like linguistics textbooks. In theory, the models could simply remember and re-store the information they were given during training.
To avoid this, Beguš and his colleagues created a linguistic test in four parts. Three of the four sections involved asking the model to analyze specially constructed sentences using tree diagrams, first introduced in Chomsky’s seminal 1957 book. Syntactic structures. These diagrams divide sentences into noun phrases and verb phrases and then divide them into nouns, verbs, adjectives, adverbs, prepositions, conjunctions, etc.
One section of the test focused on recursion—the ability to embed expressions in expressions. “The sky is blue” is a simple English sentence. “Jane said the sky was blue” embeds the main sentence in a slightly more complex sentence. Importantly, this recursive process can go on forever: “Maria wondered if Sam knew that Omar had heard Jane say the sky was blue” is also a grammatically correct, if awkward, recursive sentence.
