Syntax: A cognitive approach

One of the most celebrated features of human language is its generative power: from a restricted vocabulary of words, we can construct a vast number of new sentences, to express diverse meanings. This generative capacity is underpinned by syntax or grammar—a set of rules for how words go together. The syntax of human languages has long been argued to be complex, and even unlearnable from the input alone (Chomsky’s Universal Grammar proposal). However, the success of large language models (LLMs) has challenged this idea.  In this talk, I argue for a simple view of syntax, where the syntax of a language is just the set of dependency rules, and the dependencies are dictated by how words depend on one another for meaning. Conceptually, this simple approach obviates much of the complex machinery that Chomsky and colleagues have long postulated, including eliminating the learnability problem. Empirically, it accounts for diverse phenomena in human language processing, where non-local dependencies are costly, and explains cross-linguistic word order universals, which stem from the tendency to minimize dependency lengths.  I also discuss one interesting exception to this tendency for dependency-length minimization, which is legal language (or legalese) and speculate on possible reasons for the high prevalence of long-distance dependencies in legalese. Finally, I speculate that LLMs, similar to human children, learn the dependency grammar from linguistic patterns, leading to their impressive syntactic competence.

Link to session