Learnability

by

Geoffrey K. Pullum

• LAST REVIEWED: 30 August 2022
• LAST MODIFIED: 28 October 2011
• DOI: 10.1093/obo/9780199772810-0077

Introduction

The study of language learnability is not the same thing as the study of how language acquisition actually proceeds, either in infants or among mature speakers. Learnability deals with how any agent at all—human, animal, or machine—could in principle achieve such a thing as acquiring a language. The kinds of questions raised include what input data are needed and what procedures would work for acquiring mastery of what sorts of linguistic systems. To a considerable extent, therefore, works on language learnability have a logical or mathematical character, not an empirical character. Related topics include the psychological and computational study of pattern recognition in psychology and the philosophical study of induction. The published works on learnability often present theorems couched in terms of abstract devices, such as Turing machines or other automata. They do not typically present empirical findings, but they do set limits on viable hypotheses and point out obstacles standing in the way of empirical theorizing about language acquisition. “Acquirability” might have been a better term than “learnability,” because many linguists believe language is not entirely learned through experience with the environment in the way that many everyday skills are; it is acquired via rapid unconscious triggering of an inbuilt language-acquisition component of our brains—a component that in effect specifies most of the details of language design in advance. But “learnability” is a familiar term, and its use here should be understood as neutral regarding the empirical realities of language acquisition. Some key works on learnability of languages emanate from fields like mathematics, logic, and informatics. The annotations in this article sometimes contain parenthetical warnings that the work is from one of these technical fields rather than from within linguistics proper.

Linguistically Based Studies

The agenda was set for the modern linguistic study of learnability in chapter 1 of Chomsky 1965. Noam Chomsky not only proposed a specific form for transformational grammars but also speculated on how an already-known universal definition of them might provide the basic framework of a way a child might hit upon a correct grammar despite having relatively little to go on by way of exposure to evidence about grammatical structure. The idea was that if grammars were widely distributed in the logical space defined by compatibility with linguistic data, gleaning relatively few scraps of information about structure might guarantee that only a small number of permissible grammars were still in play as possibilities. Wexler and Culicover 1980 summarizes some of the results of work in this vein that had been achieved fifteen years later, and Gibson and Wexler 1994 and Berwick and Niyogi 1996 give a good sense of where the research program stood after about thirty years. Smolensky 1996 and Tesar and Smolensky 1998 exemplify the work on learnability to which optimality theory (OT) has given rise, and Hale and Reiss 1998 offers a direct critique of the OT approach.

• Berwick, Robert C., and Partha Niyogi. 1996. Learning from triggers. Linguistic Inquiry 27:605–622.

  A follow-up study of the system explored in Gibson and Wexler 1994 showing how to model the behavior of Berwick and Niyogi’s triggering learning algorithm on a parameter space with a Markov chain, giving some results about how and where it will work and expressing some doubts about the algorithm’s psychological plausibility.

• Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press.

  Chapter 1, especially section 8, is the locus classicus for Chomsky’s original presentation of the view that syntactic theory has something to contribute to the learnability problem. A work that absolutely every linguist should read.

• Gibson, Edward, and Kenneth Wexler. 1994. Triggers. Linguistic Inquiry 25:407–454.

  A study of how a learning algorithm could in principle discover aspects of syntax, such as the possibilities for basic constituent order, on the basis of a finite set of parameterized principles set by “triggers,” which are readily recognized expressions that provide unambiguous indications concerning the setting of some parameter. The results are somewhat pessimistic.

• Hale, Mark, and Charles Reiss. 1998. Formal and empirical arguments concerning phonological acquisition. Linguistic Inquiry 29.4: 656–683.

  DOI: 10.1162/002438998553914

  A critical response to Smolensky 1996. Taken together, the two papers provide a very useful introduction to the way learnability looks from an optimality theory perspective and what the problems might be.

• Smolensky, Paul. 1996. On the comprehension/production dilemma in child language. Linguistic Inquiry 27:720–731.

  An interesting discussion of learnability from an optimality theoretic perspective.

• Tesar, Bruce, and Paul Smolensky. 1998. Learnability in optimality theory. Linguistic Inquiry 29:229–268.

  DOI: 10.1162/002438998553734

  Detailed presentation of the constraint demotion algorithm for learning optimality theoretic grammars, in which a grammar is a ranking defined on a putatively universal set of constraints on linguistic form. Includes comments relating the approach to others, such as Gibson and Wexler 1994.

• Wexler, Kenneth, and Peter W. Culicover. 1980. Formal principles of language acquisition. Cambridge, MA: MIT Press.

  A major study arguing that E. Mark Gold’s challenge to transformational generative grammar had to be answered and an attempt to answer it with a proof that if certain stringent (but independently supported) restrictions on the power of transformations are imposed, and the radical assumption of innately known universal base is made, a transformational grammar can be identified in the limit on the basis of text input.