copyright ©2000 by William H. Calvin and Derek Bickerton

The handheld version of this book is available from  amazon.com or direct from MIT Press.

Webbed Reprint Collection

---

William H. Calvin

University of Washington
Seattle WA 98195-1800 USA

A Many-Voiced Symphonic Sentence

Well, lunchtime was hardly an appropriate setting to explain the first eight chapters of Cerebral Code. I never got to the role of corticocortical coherence in nested embedding, which is where the common code really shines. It seems capable of making syntax an everyday, subconscious task. Now that we=ve discussed Darwin Machines, coherence, and throwing=s segmented planner, it is easier to see how they all interact with the up-from-social-calculus argument structure.

The essential nature of the little role links can be seen if you try to imagine a lingua ex machina without them, merely using the segmented planner version of the Darwin Machine and parsing via boundary words (try hanging candidate phrases and clauses onto workspace trees, using simple rules such as taking the noun after a preposition and its modifiers and making a phrase out of them). Alas, you will be left with a lot of ambiguity (multiple candidate trees will remain), and not much way to resolve the empty subject and object categories of a sentence such as AJohn needs someone to work for.@ But adding role links allows ambiguity to be quickly resolved, at the lightning speeds of language comprehension and production.

The deep structure constellation for planning need not operate in serial-ordered real time, but it does need to be packed and unpacked in a speedy way, effortlessly handling never seen before combinations of words. You can see the need for this most easily when it comes to placeholders (words such as Athat@; pronouns whose referents may be in preceding sentences). If necessary, you can expand on the placeholder, providing the full name or phrase.

This is part of what is called binding in linguistics, but the need for placeholders may have originated in the inherent limitations of working memory. We often speak of Achunking@ when a particular string of words takes on an identity of its own. Most people can remember a string of about seven random digits, such as a phone number, for long enough to repeat it back (or dial the number). But they have trouble with longer strings B unless some of them form familiar chunks, such as the 1-212 string for New York or the 44-171 string for London. You also see short-form substitutes in acronyms (the way we write AVS@ as a shortcut for AVilla Serbelloni@) and internet addresses (the Rockefeller Foundation is rockfound.org). It=s really seven chunks that you can easily handle, not seven digits or words. You may use the short form at a certain level of operation but you usually need to unpack it eventually. You must be able to query the workspace where the long form is kept.

I assume that something similar is true for a sentence=s subsidiary phrases and clauses: that a short form will do for competitive purposes, so long as it can be related to the long form when the time finally arrives for the successful sentence to be Aread out@ into surface structure and speech. This suggests that messages can be sent back and forth between the Darwin Machine performing at the sentence level and the various Darwin Machines implementing the clauses and phrases. And that must require the coherence I discussed in chapter 8, so that a long-distance code could be formed up on the fly, even for seldom used prepositional phrases like Awith one black shoe.@

Assembling words into associations (Ablack shoe@) can be easily done using superpositions of hexagons, the kind that you get near borders where one hexagonal mosaic (the S for Ashoe@) overlaps with another (B for Ablack@). Form up an SB mosaic and, at its edge, superimpose it upon the code for Awith,@ and you have a prepositional phrase. Keep going and you might achieve a territory of clones, each of which contains the superimposed codes for the eight words of Athe tall blond man with one black shoe.@ It=s the Bickertonian utterance length problem in protolanguage, and Sontag=s breakfast table advice about brevity when speaking to Italian waiters in English. There are simply too many words, and you don=t know which modified what, so that you make Ablond black man@ errors.

Doing all the association via superpositions at a borderline between territories does get to be awkward B but there=s another, better way. The coherent replica of a hexagonal mosaic in a distant cortex allows, say, eight codes to be superimposed with ease (it just takes eight corticocortical paths terminating in the same area). It would seem, at first glimpse, to produce an equally ambiguous superposition. But reciprocal corticocortical links allow you to have your cake and eat it too, as they support structuring. It=s all like a fancier version of the Hallelujah Chorus.

The back projection from area alpha to area beta (on, of course, a different one-way street) can use the same code, and that means that beta can contribute to maintaining a chorus above a critical size in alpha (they are, presumably, always adapting and thereby falling silent). It would be like missing choir practice but participating anyway via a conference phone call B and perhaps making the critical difference that keeps the performance from faltering.

Alpha=s backprojected spatiotemporal pattern might not need to be fully featured, nor fully synchronized, to help out with beta=s chorus. It might be like that sing along technique called Alining out,@ where a single voice prompts the next verse in a monotone and the chorus repeats it with melodic elaboration; some singing at a fifth or an octave above the others, some with a delay, and so forth.

The backpath could also include procedural prompts, just as choirmasters and folk singers manage to include exhortations with the desired text. Procedural prompts provide one way of resolving ambiguities when decomposing embedded phrases during production by, in effect, querying an audit trail. (AWho mentioned X? Sing it again, the whole thing!@) With such structural links connecting the top-level hexagon=s spatiotemporal pattern to subsidiary ones, there=s no longer a danger that the mental model of the eight-word amalgamation Athe tall blond man with one black shoe@ will be scrambled into Athe blond black man with one tall shoe.@

The same mechanisms that clarify prepositional phrases probably help us to understand full sentences with independent clauses (AI think I saw him leave to go home@). The closed-class words, because they are limited in number, can probably all be handled as special cases, each with their own completion requirements regarding role links.

Verbs have more multifaceted completion requirements. Each verb has a characteristic set of links: some required, some optional, some prohibited. The conglomeration feels like a proper sentence only if all the obligatory links are satisfied and no words are left dangling, unsupported by a structural role. AGive@ requires three nouns with appropriate role tags, Asleep@ cannot tolerate more than a sleeper except via a prepositional phrase.

What keeps the top-level hexagon happy enough to reproduce effectively in a copying competition with other variant interpretations, such as that Ablond black man@? Presumably, a few alternatives assemble in parallel until one gains the strong Alegs@ needed to allow it to become robust enough to establish hegemony. If, in AI think I saw him leave to go home,@ the Aleave@ link stumbles, the Asaw@ hexagons might not compete very effectively, and so the top level dangles.

So the Ameaning of the sentence@ is, in this model, an abstract cerebral code whose hexagons compete for territory with those suggesting alternative interpretations. Phrases and clauses require coherent corticocortical links to contributing territories, having their own competitions and tendencies to die out if not reinforced by backprojecting codes. Weblike crosstalk between subchoruses presumably occurs, and may be quite useful so long as it remains weak enough not to show up on the audit trail.

It starts to look like a choral work of many voices, each singing a different tune but with the requirement that it mesh well with all the others. Indeed, the symphonic metaphor might be appropriate for the more complex sentences that we can generate and understand. Certainly the reverse-order analogy to Benjamin Britten=s Young Person=s Guide to the Orchestra, the all-together version being succeeded by the various voices playing separately, is the best metaphor I know for the read-out process that converts the parallel-structured plan into serial-ordered speech.

Though the common code for many cortical areas is obviously a Good Trick, is it the Good Trick that transformed RA-augmented protolanguage into fluent syntax?

One quick qualifying test is to consider the implications of efficiently linking the concept filled temporal lobe with the prepare-for-action frontal lobe, with a common code replacing the degenerate codes B and then dropping back to the old system, with incoherent paths forcing a reliance on slowly established associative links. Does it degrade gracefully, as communications engineers try to achieve with digital packet-based systems?

Without coherence, you=d still have a vocabulary (the temporal lobe still works). You=d still be able to plan some nonlanguage actions (you=d pass many of the neuropsychological tests for frontal lobe functioning), but your ability to quickly invent new trial run associations would suffer. Not only couldn=t you form up a syntactic sentence to speak (except for stock phrases), but you couldn=t judge sentences that you heard someone else speak because you could no longer judge the quality of your trial interpretations, whether they were nonsense, good guesses, or sure things. Your quality associations would be too slow for the windows of opportunity, and the results would be of poor quality because not shaped up very far by Darwinian copying competitions in the brain. And so your performance on language tasks would drop back to something like protolanguage, a wide choice of words but with novel sentences limited to just a few words to avoid ambiguity.

Another type of pathology I can imagine (chapter 11 of Cerebral Code has many more) concerns the hegemony requirements of the top-level hexagonal competition (call it the alpha mosaic). Suppose that there have to be N_action hexagons singing in the plainchant chorus before an action sequence is triggered (or, in the comprehension version of the language task, you decide the problem is solved and you can move on). Suppose that, in order to communicate with subchorus beta, it takes only N_link singers in alpha. Now suppose there is noise in the backpath, raising the critical number needed in alpha in order to keep beta singing. What happens if N_link becomes larger than N_action? Well, you couldn=t incorporate distant clauses and phrases B you could neither read them out nor have their maintenance affect the top-level competition in alpha. Only the incoherent version of the alpha code would arrive in beta and, unless it was the incoherent code for a common phrase that beta would recognize, it would be ineffective.

So degraded coherence might well cause most aspects of syntax to degrade even more abruptly. If you pull the incoherence card, it will read:

Coherence, as I briefly mentioned earlier, can also help concepts to establish colonies or branch offices in distant areas of cortex. To avoid the delays inherent in using concept resonances in distant places, you could import some of the more frequently used resonances into the frontal lobe. Nouns that started out as temporal lobe specialties could secondarily operate out of the frontal lobe, following such a metastasis.

In a two-level Hebbian memory system, the long-lasting resonances are usually produced by a sufficient number of repetitions of the spatiotemporal firing pattern. If beta=s mosaic is frequently seeding a chorus in alpha, then a resonance for it might also develop in alpha (perhaps not within the projection area of the coherent bundle, but somewhere in the territory of the mosaic secondarily generated thereabouts). In computers, this would be known as a cache (keeping frequently used code as close as possible to the processor). But it need not be a temporary resonance (as the cache analogy suggests), as it could be consolidated there (which is why I used the colony and branch office as analogies).

Furthermore, it need not be just the code for concepts; it could be code for performance, the little subroutine that it takes to quickly perform certain algorithms (the so-called Acortical reflexes@ so handy for quickly hitting the brakes if someone shouts AStop!@). The only real difference between the arbitrary spatiotemporal code for a sensation and that for a movement is whether it meshes well with the output pathways from the cerebral cortex and actually moves muscles in a coordinated way.

Within much of the cortex, code is simply anonymous code. If a spatiotemporal pattern can be copied, it may serve as a code; if it can be coherently sent over long distances, it may become a code common to many neocortical areas; if it fits the requirements of output pathways, it may even find its way out of the brain into the real world.

Other higher intellectual functions (music, planning for tomorrow, logic, playing games with rules) may, more generally, benefit from the neural systems that are so essential for syntax. Any task that requires the progressive improvement of quality would benefit from syntax=s Darwin Machine. The discovery of order amidst seeming disorder might become much easier. The segmented planner suggests a way of creating new levels of abstraction for relationships, a way to compare relationships and generate metaphors B while still being able to decompose the whole into specific actions, such as speaking a sentence. (Sometimes, we can=t get the output aspect together Bwe Aknow things of which we cannot speak.@)

Thought often has to span many possible levels of explanation and locate an appropriate one. As we try to speak usefully about a subject like language or the brain, we are often torn between dwelling on rock-solid details and speaking in terms of perhaps too abstract generalities. We need them all, but we can only speak about one at a time.