Microstructure of Cognition (Rumelhart, McClelland, & the PDP Research ... Donald Mathis, Yoshiro Miyata, Alan Prince, William Raymond, Melanie Soderstrom,.
PsycCRITIQUES March 7, 2007, Vol. 52, No. 10, Article 15 © 2007 American Psychological Association
Explaining Language Behavior: Still a Major Focus in the Field of Cognitive Science A Review of The Harmonic Mind: From Neural Computation to OptimalityTheoretic Grammar by Paul Smolensky and Géraldine Legendre Cambridge, MA: MIT Press, 2006. 1,222 pp. $100.00 (set) Volume I: Cognitive Architecture. ISBN 978-0-262-19526-3. Volume II: Linguistic and Philosophical Implications. ISBN 978-0-262-19527-0. http://dx.doi.org/10.1037/a0006558 Reviewed by
Shelia M. Kennison
Although the topic of language processing may still be viewed as a subspecialty within cognitive psychology, it has played a central role in the history of the broader field of psychology. The cognitive revolution coincided with Noam Chomsky's (1959) critical review of B. F. Skinner's (1957) Verbal Behavior. The review explained that language cannot be adequately explained via behaviorist learning principles, because most sentences that language users hear are novel combinations of words that they have likely never heard before and are unlikely to ever hear again. Chomsky further argued that children must be born with knowledge of language, specifically knowledge of the shared features of all possible human languages, and that language acquisition occurs as children discover which grammar rules among the entire set of all possible grammar rules apply to the language(s) they are learning. A reader-friendly introduction to this view can be found in Steven Pinker's (2000) The Language Instinct. Contemporary proponents of this view, including Pinker, presume that language is special—different from other types of human cognition. Over the last several decades, evidence supporting the notion of innateness of language has come from studies of inherited forms of language disorders affecting individuals with intelligence in the normal range (Rice & Warren, 2004). In addition, the view that language and other types of cognition can develop independently has been supported by the discovery of Williams syndrome, in which individuals have highly developed language ability despite severe cognitive impairment from birth (Bellugi & St. George, 2001). Since 1959, Chomsky's view has come to be just one of multiple views of language and language processing. The leading alternative to Chomsky's view rejects the notion that
language is special; rather, it is argued that language is learned via the same processes that all other cognitive skills are learned. The view shares philosophical similarities with behaviorism but differs substantially in its approach, relying primarily on simple computer models of learning whose architectures are viewed as analogous to the neuronal organization of the brain. Proponents of this view include the late Elizabeth Bates, Donald Rumelhart, and James McClelland, among many others. In 1986, Rumelhart and McClelland published their groundbreaking work, Parallel Distributed Processing: Explorations in the Microstructure of Cognition (Rumelhart, McClelland, & the PDP Research Group, 1986). The work invigorated the debate by providing tutorials in how to build computer models to investigate claims regarding how human learning occurs. Although there have been numerous investigations showing that certain aspects of language-related learning can be explained with such models (Christiansen, Chater, & Seidenberg, 1999), critics remain skeptical of such models, because the distributed nature of memory representations in these models appears irreconcilable with the notion that there are specialized modules in the mind/brain involved in language processing and separate modules involved in nonlinguistic (i.e., general cognitive) processing. The debate regarding language knowledge and language processing has invigorated the interdisciplinary field of cognitive science, which has been broadly defined as the science of the mind or of intelligence (Luger, 1994). The field is composed of scientists working in the wide range of fields, including linguistics, philosophy, anthropology, biology, experimental psychology, neuroscience, and computer science, particularly the area of artificial intelligence. Readers with interests in language and cognitive science are likely to be intrigued by Paul Smolensky and Géraldine Legendre's The Harmonic Mind: From Neural Computation to Optimality-Theoretic Grammar. One may assume from the title that the book is not intended for a general audience. This assumption would be correct. This twovolume work has been prepared specifically for those already knowledgeable of cognitive science, computational models of learning, and linguistics. In the preface, the authors recommend that the chapters in these two volumes should be consumed piecemeal, as the content in many of the 23 chapters is exceedingly technical in terms of mathematical description and theoretical impact. The authors have anticipated the difficulty that readers will experience and have provided expository boxes to assist readers in discovering the relatedness among chapters. Readers planning to consume the entire work may use the books for a seminar class, but they will likely find that two or more semesters will be needed to read and discuss the chapters and then to discuss the broader implications of the theory. The authors, Smolensky and Legendre, are professors of cognitive science at Johns Hopkins University and have been research collaborators for more than a decade. In this most recent collaboration, they attempt to explain how the key aspects of language grammars can be explained within a computational framework. In the two volumes, readers will find that the chapters have been prepared individually by Smolensky or Legendre, by the pair, or by other leading scientists coauthoring with Smolensky, Legendre, or both. The coauthors include Lisa Davidson, Paul Hagstrom, John Hale, Kristin Homer, the late Peter Jusczyk, Donald Mathis, Yoshiro Miyata, Alan Prince, William Raymond, Melanie Soderstrom, Antonella Sorace, Suzanne Stevenson, Bruce Tesar, Marina Todorova, Anne Vainikka, and Colin Wilson. Volume I is titled Cognitive Architecture and contains 12 chapters, which are divided into two parts. Part I is titled Toward a Calculus of the Mind/Brain: An Overview and contains
four chapters. Among the topics addressed in this section are the issue of how connectionist and symbolic processing can be related and the question of the kind of computation system that is needed to describe the knowledge of human languages. Part II contains the remaining eight chapters and is titled Principles of the Integrated Connectionist/Symbolic Cognitive Architecture. This section addresses the issues introduced in the first section, with much more technical detail provided. It is in this section that the authors introduce the notion of harmonic grammar, which is a connectionist-based implementation of grammatical properties referred to in the prior literature as language rules. Volume II is titled Linguistic and Philosophical Implications and contains 11 chapters, also divided into two parts. The first part is titled Optimality Theory: The Cognitive Science of Language and contains nine chapters. These chapters discuss the important universal linguistic principles related to syllables, simple clauses, and question forms. In this section the authors explicate their notions about how their computational approach can provide insight into neural and genetic levels of analysis. The second part is titled Philosophical Foundations of Cognitive Architecture and contains two chapters. These concluding chapters discuss the authors' view of the relation between connectionism and generative approaches to language, symbolic rules, and productivity of higher cognition. Readers will learn that Smolensky's and Legendre's views of language and cognition have grown out of optimality theory (OT), which was originally proposed by Alan Prince and Paul Smolensky in 1993 (Prince & Smolensky, 1993/2004). As described in Smolensky and Legendre, “OT offers a new conception of what it is that all grammars of all human languages share; it also provides a new formal characterization of exactly how they differ” (p. 1). In OT, word forms and sentences are produced by the interaction of multiple grammatical constraints. The way the conflict among constraints is resolved is certainly the most important part of the theory and is the part that is most complex. In this latest extension of OT, Smolensky and Legendre propose that the low-level aspects of cognition can be described formally as a massively interconnected connectionist network: “Higher cognition—such as abstract thought—is possible because the connectionist computer is specially organized… to perform symbolic computation: rule-governed manipulation of complex structures that are built of abstract symbols” (p. 1). The promise of this approach is that the symbolic processing referred to by Chomsky, Pinker, and others as rule application becomes an emergent property of the complex connectionist architecture. Because the low-level connectionist architecture is intended to be analogous to the neuronal level of processing in the brain, the proposals in the book have implications for future research in cognitive neuroscience. The authors go so far as to discuss an even lower level of description—the abstract genome. Without a doubt, these two volumes are ambitious. The authors tackle some of the most interesting grammatical facts discussed in prior linguistic research. In Chapter 12, the authors discuss the difficulty of explaining how a computational model can explain the ability of language users to efficiently comprehend any sentence of a language when there are an infinite number of possible expressions in a language. In Chapter 17, the authors describe their view of the nature of children's earliest representations of grammar. In Chapters 18, 19, and 21, they address the issue of universal language principles, a topic that is central to Chomsky's views about child language acquisition. In Chapter 18, the authors pose the question of whether universal syntactic properties can account for the regularities in children's earliest sentences. In Chapter 19, they explore the topic of online parsing (i.e., how language is incrementally comprehended). Chapter 21 is perhaps the
most ambitious of all the chapters; the authors ask, “Does the OT conception of universal grammar enable explicit model of how innate knowledge… could conceivably be realized in a connectionist Language Acquisition Device encoded in an abstract genome?” (p. 45). In the spirit of philosopher Karl Popper (1959), who asserted that all theories are hypotheses and all can be overthrown, the statistician George Box (1979) wrote, “All theories are wrong, but some are useful” (p. 202). Only time will tell how useful the theorizing of Smolensky, Legendre, and their colleagues will be. Because of the technical complexity of the theory and the relatively small number of scholars with the necessary combination of expertise (i.e., expertise in computational models and expertise in relevant linguistic theory) to develop definitive tests of the theory, it is likely to be a very long time before the theory has been sufficiently tested. Until then, the authors and their colleagues are likely to be producing their own confirmatory evidence designed to convince us that they have found the way to reconcile two heretofore irreconcilable theories of language.
References Bellugi, U., & St. George, M. (2001). Journey from cognition to brain to gene: Perspectives from Williams syndrome. Cambridge, MA: MIT Press.
PsycINFO →
Box, G. E. P. (1979). Robustness in the strategy of scientific model building. In R. L. Launer & G. N. Wilkinson (Eds.), Robustness in statistics (pp. 201–236). New York: Academic Press. Chomsky, N. (1959). A review of B. F. Skinner's Verbal Behavior. Language, 35, 26–58. Christiansen, M. H., Chater, N., & Seidenberg, M. S. (1999). Connectionist models of human language processing: Progress and prospects[Special issue]. Cognitive Science, 23, 415–634. Luger, G. (1994). Cognitive science: The science of intelligent systems. San Diego, CA: Academic Press.
PsycINFO →
Pinker, S. (2000). The language instinct. New York: HarperCollins. Popper, K. R. (1959). The logic of scientific discovery. London: Hutchinson.
PsycINFO →
Prince, A., & Smolensky, P. (2004). Optimality theory: Constraint interaction in generative grammarLondon: Blackwell. (Original work published 1993) Rice, M. L., & Warren, S. F. (Eds.). (2004). Developmental language disorders: From phenotypes to etiologies. Mahwah, NJ: Erlbaum.
PsycINFO →
Rumelhart, D. E., McClelland, J. L., & PDP Research Group . (1986). Parallel distributed processing: Explorations in the microstructure of cognition (Vols. 1–2). Cambridge, MA: MIT Press. Skinner, B. F. (1957). Verbal behavior. New York: Appleton-Century-Crofts.
PsycINFO →