Go to Semiotic Review of Books Home Page
Go to SRB Highlights
Go to SRB Archives
This review appeared in Volume 3 (2) of The Semiotic Review of Books.
Knowledge Representation and Symbol of the Mind. by René J. Jorna. Tübingen, Germany: Stauffenburg-Verlag. 1990. Pp. 237 ISBN 3-923721-75-7.
This readable, relatively short book is based on René Jorna's doctoral thesis carved out at the Institute for Experimental Psychology at the University of Groningen in the Netherlands. Jorna's main achievement is to propose a set of five criteria against which various cognitive theories can be evaluated. The hope is that through use of these criteria, the confusing plethora of idiosyncratic terms and concepts that permeate cognitive theories can be marshalled into some sort of order and that the various theories can then be discussed within a common conceptual framework. This should allow them to be compared with one another on an equal footing. Moreover inconsistencies and confusions within any given theory will become obvious when the criteria are applied. Discussing a theory in terms of these criteria should thus help to refine and polish the theory. Finally, Jorna suggests that the criteria themselves can be useful in delineating attributes of an ideal cognitive theory. This should constrain the kinds of theories that are proposed and investigated, and should provide a target for researchers in the field to aim at. The book is more than just an examination of the criteria, however. In justifying the criteria, Jorna provides a nice overview of six cognitive theories. Also discussed are six general goals that should be meet with cognitive theories, with SOAR (Laird, Newell, and Rosenblum, 1987) being presented as an example of a grand unified theory that has many of the characteristics of an ideal theory.
The book is organized into three parts. In Part I some general cognitive psychology background is provided, and the criteria are proposed. In Part I of the book six representative cognitive theories are briefly overviewed and then evaluated according to the criteria. Finally, the book concludes in Part III with an overall examination of the criteria, how useful they are, what attributes an ideal cognitive theory should have according to the criteria, and finally a discussion of unified theories of cognition.
Before proceeding to an evaluation of how well Jorna's book meets its objectives, it is worthwhile to look into some of the details of the book. Jorna first presents and justifies the criteria. The first criterion (C1) is concerned with the levels of description used in cognitive theory. Based principally on Dennet's (l978) distinctions, Jorna discusses three main levels of description: physical (or neuro-physiological or biological), functional, and intentional (or behavioural). Sometimes a theory will be concerned with only one of these levels, and sometimes more than one. Sometimes the levels are confused with one another. Jorna has selected certain of these possibilities and labelled them as valences. The four valences for C1 are that (1)the cognitive theory is mainly functional; (2) there is a confusion of physical and functional levels; (3) there is a confusion of functional and intentional levels; or (4) there is full scale confusion. Thus, the criterion has the flavour of a dimension in some theory evaluation space where cognitive theories can be mapped onto values (valances) along that dimension. This is true not only for C1, but for the other four criteria as well.
The second criteria (C2) is called by Jorna the morphological criterion. It is concerned with the kinds of symbols used in a cognitive theory, and how symbols in the representing domain relate to objects in the domain being represented. The idea of a depicting relation connecting objects in one domain to objects in another is central to this endeavour and underlies notions of whether a representation has a semantics. Issues such as the debate between procedural and declarative representations, the contrast between pictorial and symbolic representations, the distinction between presentation and representation, and the equivalences between different types of representations are discussed. There are five valences for theories along dimension C2: (l)elements in the representing and represented domain have been defined; (2) Representation is a descriptive scheme (symbol set); (3) representation is a procedure; (4) representation is a depiction relation; and (5) similarity between represented and representing domain is considered to be or not to be a necessary condition of representation.
Criterion C3 is also concerned with basic properties of representation, in this case the mapping between represented domain and representing domain. These mappings can be one, two, or three place predicates, so C3 is somewhat awkwardly called the criterion of n-place predicates. A representation can be a one-place predicate in that, like a unicorn in our world, it doesn't necessarily refer to anything in the represented domain. This is valence 1 for criterion C3. A representation can also be a twoplace predicate, normally a representation of something (e.g. a photo represents a particular person), but sometimes a representation for some agent (e.g. this memory has meaning for me). These variants form two versions of valence 2 for C3 called 2 and 2b. Finally the mapping can be a three-place predicate, being of something for somebody (e.g. the image of my old car is very vivid for me), similar to Peirce's (1931-1935) notion of the interpretant, object, and quality of a sign. A cognitive theory in which representation is a three-place predicate has valence 3 for criterion 3.
Criterion C4 is also strongly related to criterion C3 and C2 is concerned with what Jorna calls the route of reference. There are three valences here. Valence 1 is called denotation, and refers to the property of representation where some symbol in the representing domain is really meant to denote some entity in the represented domain (e.g. "snow" denotes the white substance that lies on the ground in winter). Valence 2 is exemplification, where a symbol or symbols exemplify characteristics of an entity without denying them (eg. as a carpet sample booklet exemplifies the textures, colours, and fabrics of a carpet). Finally, valence 3 is expression, where like valence 2 there is no denotation, but unlike valence 2 the reference is non-literal. Where the carpet sample booklet exemplifies real features, a particular painting may be an expression of some abstract entity like the cruelty of war.
Criterion C5 is the criterion of the features of symbols. Based strongly on the ideas of Goodman (1981), Jorna identifies three valences which a cognitive theory may take in relation to C5. A cognitive theory takes the first valence if only a symbol set is defined, but that's all. The theory takes the second valence if the symbol set also has a defined syntax. On Jorna's terms it is a notational scheme that allows terms to be syntactically disjoint and entirely differentiated (i.e. the symbols are not continuous). The cognitive theory has valence 3 if in addition to having a defined syntax, the symbol set also has a defined semantics, ie. that it is unambiguous, non-redundant (semantically disjoint) and semantically finitely differentiable. In Jorna's terms fulfilling these semantic requirements means that a symbol set forms a notational system. It is important to realize that Jorna uses these valences not to refer to the symbols in which a cognitive theory is formed, but to refer "to the symbols of which mental representations are thought to exist" (italics Jorna's p.52).Once these criteria have been discussed and valences presented for each one, Jorna then goes on to overview six typical cognitive theories. There are Marr's (1982) theory of visual perception; Kosslyn's (1980) theory of mental images; Tulving's (1983) approach to episodic memory; Anderson's (1983) production-rule based ACT theory of cognition; Kintsch's (1974) propositonal representation scheme; and Schank and Abelson's (1977) script-based approach to representation. These theories have been chosen both to illustrate a range of theories from low-level to high-level cognition, as well as to illustrate a wide variety of approaches. Jorna provides a nice succinct summary of each theory. After each summary, the five criteria are applied to the theory in an attempt to clarify aspects of the theory, to allow its strengths and weaknesses to be fully understood in the light of the general issues of representation for which the criteria stands, and to allow its contributions to be understood relative to the other theories.
The book concludes with three chapters in which Jorna tries to draw some global conclusions. In the First of these chapters the valences for each of the theories are summarized, bringing all the theories together in one place for the first time. This allows direct theory to theory comparison. In the next chapter some general issues of representation are clarified. Jorna then discusses, in terms of valences of various criteria, conditions that a computational theory of mind should meet. Jorna has actually previewed this discussion earlier, and it may be worthwhile quoting this earlier brief summary of these conditions:
A cognitive theory has to use a functional level of description (C1) and elements in the descriptive scheme have to be defined (C2) which at least form a notational scheme (C3). Three-place predicates have to be avoided, because of the danger of circularly (C3) Finally, the whole idea that mental representations would mirror external reality should be abandoned. This means, on the one hand, that resemblance between representing and represented domain is no longer a condition for representation (C2) and, on the other, that denotation in the case of a depiction relation is not necessary in cognitive psychology (C4). (p 55)
This last point is certainly controversial, especially for those espousing the need for a strong semantics for a representation scheme. Jorna's concluding chapter examines Laird, Newell, and Rosenblum's (1987) SOAR theory from the point of view of how well it adheres to the ideal conditions summarized above. Not to telegraph the conclusions, but SOAR comes out rather well!
Overall, the book is quite readable. It is not a huge, impenetrable tome written in obscure technical language. Instead, it puts forward its arguments coherently, with appropriate examples. It is also a fair book. Due credit is given to the many different points of view that have been espoused on representation and cognition, and Jorna attempts to help the reader understand and integrate these points of view. The book is not for rank novices. It has been written for those who have thought about these issues. However, it is general enough that researchers of all backgrounds, including computer scientists, cognitive scientists, psychologists, semioticians and philosophers, should all find it, in the main, comprehensible. The book has two main uses. First, it is a nice overview of six (or seven if SOAR is included) cognitive theories. The second valuable aspect of the book is the issue-centered discussion generated when Jorna presents the five criteria and their valences. The latter should stimulate debate in the field about the goals that a cognifive theory should meet.
So, at a general level the book is both useful and stimulating. However, at a more detailed level I would like to raise a few points, both about the form of the book and about its content, in particular the criteria, the valences, and the usefulness of the criteria in evaluating, comparing, and proposing cognitive theories. This discussion is not meant to shoot down Jorna's work by any means, but it does shed some light on the anxieties a reader may face as this book is read.
While Jorna's book is readable, if suffers from two relatively minor problems of form. First, it has been somewhat carelessly proofread. Figures have been mislabelled (e.g. Figure 4.2 is obviously meant to be Figure 4.1). An entire column in the table of Figure 9.3 has been left out, even though subsequent discussion assumes the column is there. There are many examples of careless editing (e.g. "which Schank & Abelson elaborate by Schank & Abelson in great detail" - 168). The book has clearly been "spell checked", but this doesn't prevent the frequent choice of an entirely wrong word (e.g. "short time memory" - p.149). Some of these problems may be related to the other minor difficulty: the book in its original thesis form was written in Dutch and has had to be translated into English. This sometimes results in awkward English expressions, although it must be emphasized that overall the book uses English very competently.
Next, I would like to move on to matters of content. To start I would like to evaluate the criteria themselves. The criteria are explicitly non-orthogonal. This is especially so for criteria C2, C3 and C4, which all involve aspects related to issues of semantics. The interdependence among the criteria can be confusing, a confusion that is compounded by the somewhat wandering philosophical discussion Jorna embarks upon as each criterion is presented. While this discussion is necessary if a balanced perspective is to be achieved, it does tend to obscure the truly central aspects of each criterion. It might have been preferable to more straightforwardly present the core ideas underlying each criterion, and save the philosophical discussions until sometime later in the book, perhaps in a summary chapter in Part III.
The valences are also somewhat problematic. A valence is achieved by mapping a cognitive theory onto a particular value associated with one of the five criteria. Sometimes this valence is a primitive feature of the theory, as in C5 where each valence directly describes a property of the symbols used by the theory. On other occasions, this valence is at one remove and is an analysis of how the theory maps onto properties of the theory, rather than being a direct indicator of the properties themselves. This occurs in criterion C1 where the valences are not primitive indicators of whether the theory is physical, functional, or intentional, but are instead an evaluation of how confused or not the theory is about the physical, functional, and intentional levels. The schizophrenic nature of valences is certainly confusing; in fact, the valences are not used consistently even by Jorna. For example in summarizing the theories in part III, he presents a table (Figure 9.1) outlining the valences for each theory on criterion C1. However, this table does not assign to each theory one of the four valences of C1, but instead indicates which of the three primitive indicators (physical, functional, or intentional) the theory has. Thus, even Jorna has used the valences for C1 inappropriately.
Criterion C1 also illustrates another problem with the valences: they sometimes do not cover all the possibilities. What if a cognitive theory is mainly neurological (as in some of the more recent connectionist architectures)? There is no valence to describe this state of affairs, unless the theory also confuses the neurological (physical) level with the functional level, in which case valence 2 might apply. If the valences for C1 were re-formulated to be the primitive properties physical, functional, and intentional, then this connectionist theory could be said to be physical. Other more confused theories could be said to have two valences, say physical and functional and these valences could be accompanied by a metastatement that the valences are confused. This solution would help to get rid of gaps in the valences, and would also get rid of the schizophrenia in how the valences are used that was mentioned above. It would, of course, mean that a theory could take on more than one valence on a given criterion, a situation which Jorna may have been trying to avoid. If this were the objective, however, it seems that even now this goal has not been achieved. Take criterion C2 which has five valences, several of which can be simultaneously applied to a theory. Many logic-based theories, for example, have defined elements in both representing and represented domains (valence 1) and also view representation as a depiction relation (valence 4).
A final problem with both the criteria and valences is their failure to capture every aspect of a cognitive theory. Of course, to expect total coverage is unrealistic, but Jorna falls considerably short in coverage. Jorna often finds it necessary to go outside of the valences and even sometimes the criteria in order to critically evaluate some aspect of a theory. For example, Goodman's (1981) distinction between autographic representations (where history and intention play a role) and allographic representations (where they don't) is frequently used throughout Part II, especially in discussing how a theory meets criterion C5. In fact, this distinction often seems to be more central than the valences actually associated with C5. Such frequent dependence on concepts external to Jorna's categorizations tends to diminish the importance of the criteria. The reader is also left in the awkward postion of having to constantly go back and reread aspects of Jorna's discussion that when first read were assumed to be secondary.
I would now like to move on to general matters relating to both form and content. The first of these is a difficulty faced by any approach that defines a new set of terms and then must rely on the reader to understand them deeply in order to read the rest of the book. I call this difficulty the "chicken and egg" problem. In Jorna's book this problem arises as follows: understanding the criteria is a crucial prerequisite to understanding Jorna's use of the criteria as the various criteria theories are evaluated, but the criteria themselves cannot be easily understood until after the reader has seen them in use during the evaluation. This chicken and egg cycle can only be broken by flipping back from cognitive theory evaluations in Part II to first definitions in Part I in order to resurrect the context surrounding the critera and valences. The problem becomes a diminishing one as the reader progresses through the theories, and in any event seems unavoidable when a multi-dimensional discussion has to be mapped onto a linear sequence. The lesson is that, while the book is readable it also requires the effort any subtle and interesting technical presentation normally requires.
Another danger arises whenever as in this book, categories are created in order to explain phenomena. This is the "square peg round hole" phenomenon. When evaluating a theory, Jorna must interpret the theory in terms of the criteria. This can lead to great insight when the mapping points out mistakes in the theory or fundamental similarities between the theory and some other theory. However sometimes the mapping is not a straightforward process, and aspects of the theory must be considerably distorted in order for the transformation to be accomplished. The theory's square pegs must be fit into the criteria's round holes. This can lead to conclusions being drawn about a theory which are not really conclusions about the theory itself but are conclusions about the distorted theory that has resulted from the transformation of the theory into the criteria. The theory is not considered on its own terms. Thus, when Jorna criticizes certain aspects of a theory, there is always the residual doubt that what is really being critiqued is some perturbed version of the theory which the original creators of the theory might very well reject. Perhaps the criticism should be of the criteria, or the mapping that has been done from the theory to the criteria, not of the theory itself.
For example, in the discussion of Kosslyn's (1980) theory of mental imagery Jorna claims "the debate should not be about propositions versus images or abstract versus concrete ... but it should be about the direct properties of symbol sets which lead to different representations "(p.79). Depending on how one views notions such as propositions, images, etc., this is either a great insight into what heretofore has been a confused issue, or it is a misinterpretation of Kosslyn's terminology. It depends on whether or not you agree with Jorna's mapping of these issues onto issues involving symbol sets. From my point of view, there do not seem to be any outrageous examples of the "square peg round hole" phenomenon in Jorna's exposition, although the arguments are often quite subtle and other readers may disagree.
In any event Jorna has gone to some extent to some length to minimize this problem. In the exposition of each theory in Part II, Jorna carefully divides the discussion into a criteria independent elaboraton of the theory more or less on its own terms followed by a mapping of the theory onto the criteria. Thus, the square peg round hole phenomenon is isolated to only the last section of each chapter. This approach also has the benefit of allowing readers who do not want to invest the effort needed to understand the criteria to simply access the synopses of each theory contained at the beginning of each chapter and ignore the criteria-oriented discussion at the end.
There is one final danger inherent in an approach that attempts to clarify and rationalize diverse notations by inventing a new, unifying notation. The danger is that the new notation will add still more terms to the existing babble, and confuse things even more. Jorna has tried to minimize this problem, to some extent anyway, by adapting existing terms from people like Goodman, Dennett and others. Rather than devising a wholly new language, Jorna has created a sort of Esperanto of the philosophical psychological, semiotic, and computer science literature, borrowing from here and there as necessary. But there is a danger in this too. The new language may not match the needs of any particular interest group well enough that they are wiling to give up their own terminology, and thus, like Esperanto, it will ultimately prove to be a noble, but failed experiment. Only time will tell whether Jorna's approach will gain widespread acceptance.
In summary I would like to commend Jorna for tackling a difficult task. Any attempt to rationalize the plethora of competing theories and complex issues surrounding representation and theory creation in cognitive science is bound to be somewhat controversial. Jorna has tried with commendable objectivity to provide a conceptual framework for understanding competing cognitive theories. To a very great extent Jorna has succeeded. I would certainly recommend this book to anybody with deep interest in issues of representation and cognition, whatever their background.
Anderson, John R. (1983) The Architecture of Cognition. Cambridge, Massachusetts: Harvard University Press.
Dennett, Daniel C. (1978) Brainstorms. Brighton, Sussex: The Harvester Press.
Kintch, Walter (1974). The Representation of Meaning in Memory. Hillsdale, New Jersey: Erlbaum Associates.
Kosslyn, Stephen (1980) Image and Mind. Cambridge, Massachusetts: Harvard University Press.
Laird, J.E., Newell A., and Rosenblum, P. (1987) "SOAR: An architecture for general intelligence." Artificial Intelligence 33.1:1-64.
Marc, David (1982) Vision. New York, New York: Freeman and Company.
Peirce, Charles Sanders (1931-1935) The Collected Papers of Charles Sanders Pierce Volumes I-IV. Cambridge, Massachusetts: Harvard University Press.
Schank, Roger C. and Abelson, Roger (1977). Scripts, Plans, Goals, and Understanding. Hillsdale, New Jersey: Erlbaum Associates.
Tulving, E. (1983) Elements of Episodic Memory. Oxford, England: Oxford University Press.Gordon McCalla is a Professor in and Head of the Department of Computational Science at the University of Saskatchewan in Saskatoon, Canada. His research interests span a wide variety of topics in artificial intelligence (AI) including knowledge representation, natural language dialogue, and planning. Currently, his research is focused on how artificial intelligence affects and is affected by the attempt to incorporate AI techniques into intelligent tutoring systems (ITS). Professor McCalla helped to co-found the Laboratory for Advanced Research in Intelligent Educational Systems (ARIES) which is now Canada's leading centre for research into intelligent tutoring systems. He is author of over 60 articles on many aspects of AI and ITS, serves on the editorial boards of three journals (including the journal Computational Intelligence of which he is coeditor ), and has coedited with Nick Cercone a collection of essays on various aspects of knowledge representation entitled The Knowledge Frontier, published by Springer-Verlag in 1987.