Formal principles of language acquisition By Kenneth Wexler and Peter W. Culicover (review)

REVIEW ARTICLE Formal principles of language acquisition. By Kenneth Wexler and Peter W. Culicover. Cambridge, MA: MIT Press, 1980. Pp. xvii, 647. $35.00. Reviewed by C. L. Baker, University of Texas, Austin* This book can best be viewed as an interim report on a long-term research project at the University of California, Irvine. The aim of this project has been to develop a satisfactory theory of language learning—one which incorporates an adequate theory of generative grammar, and also a learning procedure that can pick out a descriptively adequate grammar for a language on the basis of a finite sample of data. While investigators at several other universities have been interested in this general problem for varying periods of time, this book is by far the most systematic and careful study that has appeared to date. Earlier work on this project led to the preliminary result (reported in Hamburger & Wexler 1975) that, given certain assumptions about input data and about the form and functioning of transformational grammars, it was possible for a learning procedure to attain a correct transformational grammar for a language on the basis of a sample of the language in question.1 The authors of that paper made two background assumptions about the nature of the problem under investigation. First, they assumed that each primary datum provided by the learner's environment consisted of a base phrase-marker paired with a surface string. The rationale for assuming that the language learner was provided with information about the base phrase-marker rested on the assumption that this phrase-marker would be close to the semantic representation of the incoming utterance, and that this semantic representation could be deduced by the learner from the situation in which the utterance occurred. A second basic assumption was that the learner's central task in the acquisition of a generative grammar was the construction of a transformational component. Both these background assumptions are carried over into the present work. What sets the newer study apart from the older one is the concerted effort to show not only that transformational components can be learned, but also that they can be learned on the basis of exposure to relatively simple primary data. A major deficiency of the 1975 theory was that the constraints it assumed concerning the form and functioning of transformational components were relatively weak—with the result that components were allowed which could require, for their acquisition, sentences containing several thousand embeddings .2 The newer, more restrictive theory proposed in the present work makes possible a drastic reduction in the degree of complexity required in the primary * Work on this review was supported by the National Science Foundation through Grant BNS7924672 to the University of Texas, Austin. 1 More precisely, the probability of arriving at a correct grammar could be made arbitrarily close to unity by choosing a sufficiently large number of input data. 2 The complexity of primary data required for successful learning was a function of certain properties of the base component. 413 414LANGUAGE, VOLUME 58, NUMBER 2 (1982) data: now a correct transformational component can be selected on the basis of sentences that contain no more than two embeddings. To use the terminology of the book, the central result is that degree-2 learnability is possible for the class of transformational components whose form and functioning are as specified. W&C's book is divided into seven chapters. Chap. 1, 'Methodological considerations ', is primarily concerned with clarifying the concept of language acquisition to be assumed in the work. The central point is that the notion adopted is much closer to that of Chomsky 1965 or Gold 1967 than it is to that assumed in most developmental studies. Chap. 2, 'Foundations of a theory of learnability', introduces the background assumptions about the task at hand, spelling out the definition of learnability to be used and the type of input data to be assumed. Chap. 3, ? learnability result for transformational grammars', reviews the Hamburger & Wexler paper, with primary emphasis on the learning procedure employed there. This learning procedure is the part of the earlier theory that has been carried over to the present work. Chap. 4, 'Degree-2 learnability', is the central...