Abstract

Language has rules and rules have exceptions: think the English regular verbs that all add “-ed” and the some 150 irregular verbs that don’t. Children learning languages, as well as statistical models of natural language, must be able to accurately determine the scope of linguistic patterns.

In this talk, I discuss the Tolerance Principle, an empirically motivated cost-benefit calculus for assessing the productivity of rules and exceptions in human language. The Principle is stated as a simple parameter-free equation (hint: it’s not Bayesian), and is intended to be a psychological model of language learning but can also be effectively applied to NLP tasks. I review evidence from child morphological and syntactic acquisition, including results from artificial language learning experiments with young children. I will also discuss how rule learning can be applied to counting and numerical development in children, revealing some potentially deep connections between language and number.

(The artificial language learning studies are joint work with Kathryn Schuler and Elissa Newport at Georgetown University.)

Biography

Charles Yang teaches linguistics and computer at the University of Pennsylvania and directs the Program in Cognitive Science. He works on computational models of language learning, variation, and change, and is the author of four books, including The Price of Linguistic Productivity: How Children Learn to Break the Rules of Language (2016, MIT Press).