The Architecture of the Language Faculty
Ray Jackendoff, 1997

This is definitely an interesting book.

Uncommitted Working Space
It is often assumed that brain operations, such as the construction of a novel sentence, must require uncommitted working memory as data space. See, for example, (p 8), or Calvin's
The Cerebral Code . This is not necessarily the case. If, for example, the brain works by forming interconnecting links between arbitrary subsets of predefined representational elements, then the working data pattern can be fully represented in the temporary interconnections. No additional, uncommitted neural structures are required. There is some evidence for reentrant structures which could support such links. Such interconnecting links, of course, also form the core of Edelman's Neural Darwinism theory of brain function.

The above argument is not quite correct. J. says uncommitted memory is required, but does not say that this implies uncommitted neurons. Calvin's theory, on the other hand, explicity requires uncommitted neurons.

How Does the Brain Do It?
While repeating that the brain must somehow manage language processing, Jackendoff laments that precious little is known about how it works. It is my hope that this web site will help provide the connections between many fields of inquiry that will be required to meet Jackendoff's lament. Given that his words were written as long ago as 1987, I believe that we may already be much closer than the pessimistic view he outlines on p 9.

One place I disagree (in emphasis only) with Jackendoff is that I would carry the banner of repealing syntactocentrism a little farther than he has. I would argue that perceptual and production pathways are distinct, coming together only at the conceptual level. I believe lexical entries include both perceptual and production sections, more or less separate and coded in relatively direct terms of auditory decoding and motor control structure, respectively. This, of course, relegates much of current linguistic theory to be no more than a descriptive framework. In saying that, I do not deny that during the last 30 years the linguistics community has completely and thoroughly overhauled the knowledge of linguistic structures. It is only on the basis of this overhaul that we are now beginning to be able to piece together just what it is that the brain does.

Enriched Composition
Jackendoff has presented
a number of cases in which the meaning of a sentence cannot be obtained from either the syntactic structure or the semantic constructs of the words used (or at least that to so requires an unacceptable expansion of the semantics of apparently innocuous bystanding words). I agree with his basic position that the meaning cannot therefor be represented in the Logical Form, but I would further weaken the role of syntax in the whole affair. In Section 3.6 , he is still clearly sticking to the interpretation that this is a language question.

I would argue that the job of syntax is merely to inform the listener of some of the relationships between the objects referred to by the words. I applaud Jackendoff's bravery in forging ahead and trying to pin down some of the things that must be going on in the conceptual structure. It sees to me, however, that the problem is most appropriately framed in terms of the construction of a number of possible interlinks between triggered lexical items (guided by the syntactic patterns), and followed by a sort of optimization or best fit solution to find the pattern of meaning which makes the most sense given the words, the pragmatics and anything else available to the listener. Such a process could also be viewed as a competition of the sort that Calvin discusses in The Cerebral Code . I would not expect to be able to isolate any reasonable number of statable rules to elucidate what happens in this panoply of situations.

Modular Representation Goes the Wrong Way?
The entire soul of this book, as stated in the introductory chapter and even in the title, is concerned with how the brain does it. I believe that Jackendoff's
Representational Modularity hypothesis (RMH) goes exactly the opposite direction.

There is no reason to believe that the brain isolates things so cleanly. In fact, as Jackendoff notes in chapter ?, things are downright sloppy in there. As one example, the RMH would have us believe that phonetic perception proceeds to the point of a proper phonemic analysis and only then is it allowed a look at the lexicon. My ten years working with speech recognition software tells me that, without the faintest doubt, the perception process has immediate access to the lexicon. The /t/ in the word "take" is well known to be quite different from the /t/ in the word "stake". Why should this difference be thrown away rather than being used to help locate the desired lexical item?

It is my belief that phonetic parsing uses some sort of phonetic space, perhaps akin to that described by Miller many years ago. In that space, allophones, or more generally, just phones, are tracked as individual trajectories, more or less as they are extracted from the cochlear inputs. (There are a couple of well-known speaker and language adjustments which presumably apply before the trajectory is followed in the phonetic space.) Links would be constructed from the phonetic space directly into lexical access parameters.

Now, admittedly, this entire process could be viewed as a preliminary to the truly linguistic part of the task. But, on the other hand, we don't want to get carried away with syntactocentrism.

Is There One Syntactic System that Works Both Ways?
All of this really brings up a deeper question of just exactly what is the role of a linguist in all of this? Is linguistics really concerned with the brain, or should that be left to the psycholinguists?

Linguistics, as remolded during the last 30 years by the generativists, has been faced throughout with a sort of crisis of identity. On the one hand, they wanted to get away from the non-functionality of "Descriptive Linguistics". Unquestionably, great strides have been made in understanding much of the underlying structure of language. But to view a rule as a bidirectional, information-isolated description of underlying structure is certainly not the same thing as viewing the same rule as a procedure which the brain must follow in order to exhibit linguistic behavior.

Jackendoff's blackboard model helps a lot to clarify how a rule system could serve both perception and production and still appear to non-directional. However, it still seems to me that there is a problem, such as with the perception of regular affixes. I maintain that it is essential for the phonetic perception processes to have immediate access to fully inflected forms to mediate the perception of these words. On the surface, it seems correct to reject the idea of active generation. As Jackendoff puts it, "It makes little sense to think of randomly generating a phonology, a syntax and a meaning, and then seeing if you can match them up:". However, it may be possible that certain proposed brain structures make it possible, given certain minimum cues, to implement a limited form of active generation adequate to serve the needs of the perceptual system. For exmaple, suffixes could be handled by triggering lexical access as soon as a potential word stem has been identified, making available the inflected forms, perhaps only in the form of distribution probabilities, but suffucient to allow the phonetic perceptual system to complete the identification. Whether this mechanism is feasible may depend on the (non)existence of irregular inflectional prefixes, which would disallow such lexical access and greatly increase the difficulty of the perceptual process.

Brain Stratgies for Parsing
If Jackendoff is really interested in psychological reality, then he must consider some issues involved in parsing, assigning syntactic structures to incoming sequences of acoustic data. Some interesting benchmarks are presented in Boden's
Computer Models of Mind .

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