Semantics and Cognition
Ray Jackendoff, 1983

This Book is a Success
Jackendoff successfully clarifies a number of issues which have befuddled linguists and philosophers alike. That success is due to one powerful idea. The idea itself is not new. But our intuitions are preprogrammed to reject it. The key idea is that everything we sense, everything we think and everything we believe is inside our heads, not out there in the world.

What Is This Book?
Why should this book be categorized as "Linguistics" while Searle's
Speech Acts is categorized as "Philosophy"?

Because Jackendoff actually considers linguistic facts, such as sentence structure, in his arguments. Searle basically argues in a vacuum.

This book was published 12 years before the author's Languages of the Mind , and contains precursors to the theory of conceptual structures as presented there. "Languages", however, is no mere repeat of this material. There is much worthy of serious consideration in both books.

Perception vs. The Real World
It's really refreshing to see, at last, a straight-forward
discussion of the often-repeated visual illusion figures, with the "Gee Whiz!" factor factored out. The style carries right through "[we're not at liberty to dismiss these ... as mere ... tricks ...]" to the conclusion that the entire page (or computer screen) that you are reading is really all in your head.

Jackendoff's example of "Where is Beethoven's Fifth Symphony?" (p 27) would be well pondered by more than one philosopher who has wallowed in the issue of where one might locate the mind.

A Single Unified Language/Cognitive System?
In contrast to the model presented here,
Bickerton argues that there must be two separate representational systems, one directly connected to the sensory input and one at a higher level, subject to inputs from memory (compare Jackendoff's figure 1.3).

What evidence might we examine to settle this question?

All Senses -- Not just Visual
It is clear that all sensory systems must be considered in determining when an #entity# is perceived. Although Jackendoff stresses the visual system, the point of reference must occur after information from all senses has been integrated. In other words, the question, "What's that?", can be asked while sniffing the air, just as well as when pointing to something. Perhaps this was meant to be included in the "etc." pointing to the "conceptual structures" modules in figures 1.3 and 1.4.

Seven Ontological Categories
Jackendoff makes a strong point about the
relationship between possible perceptual #entities# and the grammatically distinct structures that can be used to reply to various forms of wh-questions . His (not quite clearly stated) claim is that these categories are innate, that is, that we can use these to communicate because we all possess the perceptual mechanisms to isolate these particular #entities# from the ongoing stream of information in the sensory field.

Perhaps an example of quantification of the [AMOUNT] category might be, "Bill went {part way / all the way} to the store", although the existential and universal properties are strained a bit here.

Not Seven Categories?
An alternative picture is that not all of the above categories are innate -- that, in fact, only a small set, perhaps 3 or 4 categories are innate (the same ones used for time?) and that others are derived from the basic group. If this is so, it should show up in at least two ways. First, there should be a clear difference in the time it takes a child to learn to make sentences with the "extended" categories. and second, there should be more variablility in the use of the "extended" categories in various languages. Jackendoff (p ) mentions some variablility in language use, but does not say enough to pin down the issue.

Are There More Than Seven Categories?
Jackendoff suggests an additional category of #sounds#. However, this seems to me as no more than a different property of #thing#. Even if this particular list is not exhaustive, would it nevertheless be possible to compile an exhaustive list based on linguistic evidence? If this really did represent an exhaustive list of such categories, then what would that imply for brain mechanism research? Does a single gene or group of genes specify each category? Did they evolve all at once or each as a separate perceptual capability? Is the list complete in some real-world sense? That is, has our evolution reached a point where the world is satisfactorily represented by this particular list of categories?

Three Temporal Categories?
Pinker notes a
relationship between time and space . Three of Jackendoff's categories, [PLACE], [MANNER], and [AMOUNT], would have temporal correlates corresponding to ter Meulen's Aspectual Classes , Plugs, Filters, and Holes. These could be called something like [TIME], [PATTERN], and [DURATION].

It is interesting that the quantification of [DURATION] is strained in much the same way as noted above for [AMOUNT]. "Bill ...".

In Note 9 to Ch 3, Jackendoff mentions an "extension" pertaining to uncertain boundaries of certain kinds of #things# and #events#. Note that this is very similar to the distinction ter Meulen makes between Plugs and Holes. It is perhaps obvious why spatial structure should be so much richer that temporal structure. However, I believe these possible correspondences bear further investigation.

Mechanisms of Categorization
Jackendoff's descriptions of the effects of categorization have an appeal in a way that his implementations do not. My response to his preference rule mechanism is about the same as I react to a list of structured schema as a reasonable implementation of the lexicon. It seems to me that what is sorely needed is a physiologically plausible mechanism by which such categorization could be implemented.

Recently, a method of document retrieval has been described based on the singular value decomposition of a matrix of word occurrences vs. documents. The method is described in papers by Deerwester and Landauer (19 ) and Dumais and Landauer (19 ). In the latter paper, Dumais presents a description of the matrix decompostion method in terms of an equivalent neural network architecture. In reading that description, I was struck by the similarity of the matrix method and the descriptions of re-entrant categorizers in the Neuronal Group Selection theory of Gerald Edelman.

If this similarity is anything other than sheer coincidence, then we can see how the properties which Jackendoff imparts to the concept formation mechanisms might be implemented as a kind of neuronal similarity detector. To be sure, the exact mathematical structure of the singular value decomposition would not be implemented in a neural network. But if the exact least mean squares computation done by the SVD is replaced by a non-linear approximation as might be computed by a collection of post synaptic potentials, we have a promising mechanism by which to implement much of the perceptual processing which Jackendoff describes. The data reduction function described by Edelman as implemented by the sparcity of interconnected neurons between the two ends of the re-entrant loop corresponds quite nicely with the dimensional reduction performed directly by discarding some fraction of the diagonal coefficients of the companion matrix.

A system of such re-entrant categorizers could, very much as Edelman proposes, implement much of the extraction of conceptual figures from the perceptual data stream. In addition, a set of such pattern classifiers connected between concept structures would have the ability to extract common pattern similarities in the manner as described by Jackendoff's preference rules. Deerwester and Landauer speculate that their method of document characterization could be useful for determining word meanings.

Edelman's description of the application of such re-entrant categorizers to perceptual processes raises some implemetation issues which differ somewhat from the structures described by Jackendoff. In particular, his Chapter 7 discusses the mechanisms of preference rules as if a common set of such rules might apply across the board to most, if not all, sensory modalities and to conceptual processing as well. What is clear from the implementation Edelman describes, and as he has implemented such mechanisms in his "Darwin Machines", it is clear that a specific categorizer would be "hard-wired" in each place where it's function was needed. Although from general principles of genetic parsimony, we might expect such multiple implementations to have a lot in common, we should not be surprised if each such implementation was tailored to some extent to its particular application.


Top of Page | SC Summary | Sort by Topic | Sort by Title | Sort by Author