In an earlier paper, “Performance Constraints and Move-α,” I demonstrated conclusively, through the use of a computer model, that the current GB movement theory is inadequate to explain the actual generation of texts of spoken English.¹ Attempts to deal with the problem have been unsuccessful so far.² Indeed, it seems that there is no way to keep Move-α in its present form and still maintain the idea that performance makes use of an underlying competence whose form is dictated by UG. This does not mean, as some of my less enlightened colleagues have suggested, that we should abandon movement rules altogether. The past thirty years have clearly shown, the opinions of the uneducated notwithstanding, that no grammar can be descriptively adequate without a transformational component. Nor should we abandon the insight of more recent research, that transformations are not attuned to special structural characteristics of the categories they move—that is, there is not one separate rule which moves N'’s dominated by V’s and another that moves Whs c-commanded by empty nodes. No, all movement rules are alike in that they can move any kind of constituent to any position. They differ, however, in that each can move constituents from only one position.

In considering this hypothesis, we must realize that for too long linguists have ignored the fact that sentences are linearly ordered. The quite important realization that they are also hierarchically ordered has unfortunately caused a reaction that omits a fact about sentences that is obvious even to the layman. When we say that words are in a sentence, they come out not simultaneously, but one after another. We can identify one word as occupying the first position in the sentence, another in the second, and so on. To test the theory that movement rules apply to elements in specific positions, I constructed a computational model which would operate in such a manner.³ For convenience’s sake, I chose to name the first sentence position α, the second position β, and so on up until ω. Later, I could expand the program to deal with sentences of greater than twenty-four words in length. Thus, I had twenty-four separate non-structurally-dependent movement rules, applying linearly in modified recursive fashion to transform D-structure to one of several derived S-structures. To my delight, the program worked perfectly—given one of an exhaustive set of D-structures, it yielded all and only the appropriate S-structure, so long as the appropriate constraints were also included in the program. Furthermore, the processing time to produce a single S-structure, which is of course what we do in natural speech, was only 5.1 +/- 2.2 seconds. Variable analysis demonstrated that program refinement and transferal to a more sophisticated processor could reduce transformation time to 1.3 + 0.7 seconds. Extrapolating from Hogarth (1987), we find prediction of corresponding processing time in human computational devices of under 0.3 seconds. Thus, as a model of competence underlying performance data, this approach worked beautifully.

There was only one problem. What about sentences of more than twenty-four words in length? It has long been a tenet of scientific grammarians that there is no limit, in competence, to the number of words there may be in a sentence—the limits which we observe are purely the result of performance constraints. Now, if there are sentences of infinite length in D-structure, then the model presented so far would require an infinite number of position labels and an infinite number of transformation rules. This presented two difficulties. First an infinite number of rules would require an infinite processing time, which would put us right back where we started. Second, and perhaps more importantly, I was out of Greek letters. Clearly it was time to reexamine the claim that limitations on sentence length result from performance constraints.

I began by looking at sentences in English, not written English, nor recitation from written texts, but actual spoken English as it occurs spontaneously. Using recognized markers of phonological sentences such as intonation patterns, I examined thousands of hours of spoken texts to find out how many words were in each sentence. To my surprise, I found that the upper limit was twenty-four. In not one text did a phonological sentence containing more than twenty-four words occur. I was astounded, but, worried lest this be a language-specific phenomenon, I next gathered data from one hundred seventy-three languages from six different continents and fourteen unrelated families. The results: the same. In short, I discovered that no human being ever utters a spoken sentence with more than twenty-four words.

Now, it is important to remember that performance constraints are not rigid and precise in the way that competence is; this is one reason why, as language scientists interested in exactness and quantifiability, we linguists stick to the latter. It is entirely inconsistent with the nature of the performance that it should impose an absolute, unvarying constraint on the number of words that may appear in a sentence. Therefore, the limit on the number of words in a sentence, the existence of which is unarguable, must be part of competence. Note, however, that we do not need a separate Sentence Length Constraint (SLC) to account for the data, as we would under other theories of grammar. The limit arises naturally as an unintended but empirically verified consequence of the Linear Positions Theory.

I now turn to other phenomena connected to the LPT and the consequent SLC. We all know that in writing, sentences may have more than twenty-four words. Yet we also know that writing is significantly different from speech in other ways, with grammaticality judgments, for example, being different for the two. In short, we should not expect all modules of the grammar of spoken language to apply to written language, and this theory is but further proof of the difference between the two.⁴

Having discovered an upper limit to the number of words in a sentence, we might wonder whether there is a lower limit. Further research is needed to clarify this point, but it seems that the lower limit is also twenty-four! Analysis of processing time in actual human subjects shows that they spend the same amount of time performing the transformational operations on sentences which appear to have only two or three words as one sentences which have twenty-four. The differences in transformational processing time for sentences of any apparent length, in fact, are well within the range of acceptable experimental error. Thus, it seems that sentences which, on the surface, have only twenty words, in fact have empty positions which are nonetheless subject to movement. By the way, the reader should note the elegant similarity between empty positions and empty categories.

As a further question, we might wonder, why twenty-four? The answer to this is plain. The numbers two and three have long been recognized as the fundamental numbers in both cognitive studies and numerology. We think in twos and threes; they are essential numbers hard-wired in our brain. Now consider the number 24. Add its digits together, and you get 6, which is 2 x 3. 2 and 4 are the two numbers on either side of 3. Divide 4 by 2 and you have 2 2’s, which, in company with the other digit, make 3 2’s. As for 24 taken as a unit, it is equal to the following:

• (2 x 3) x (2 x 2)
• 2³ x 3
• 3(2²) x 2
• (2 + 2)² + (3 x 2) + 2
• 2(3²) + (2³) - 2
• 3(2 + 2 + 2) + (2 + 2 + 2)
• (3/2) x 2^{(2 + 2)}
• (3 x 3) + (3 x 2) + 2 + 3 + 2 + 2
• (3/2^{((3 x 2) - (2 + 2))}) x 2^{(3 + 2)}.

The theory presented in this article is elegant, parsimonious, mathematical, scientific, and accounts for the data. It is neither overly powerful, nor, like the standard movement theory, overly constrained. It is supported by computational evidence, as well as by real language data. Moreover, it has made empirically verified predictions about matters other than those which it was created to deal with. In short, it is perfect. It should therefore be adopted as an integral component of the theory of grammar.

Gianlorenzo Bernini

Rome, Italy

Carvaggio M. 1986. “Movimento di alfa e chiaroscuro.” DIR 17.3 226-242.

Le Nain, L. 1985. “Mouvement de la charrette.” OIG 42.1 43-74.

Ribera, J. 1983. “Atando Serapio.” IRM 3 19-21.

Van Dyck, A. 1982. Government in various settings. Trans. by Charles Stuart. London, England: Jacobite Press, Inc.

Watteau, A. 1984. “Deux théories des transformations: le Pied Piping et le berger.” USDA 73 117-183.

² Unless, like Van Rijn, you resort to arguing that competence and performance are utterly unrelated and that linguistic data has no bearing on our models. More responsible arguments include Carracci’s suggestion that the performance mechanism has a built-in Rewrite Delimitation Device (RDD), where p=0 and s=12, as well as Poussin’s attempt to generalize Trace Theory to handle the problem. Both of these efforts fail, Carracci’s because it would make sentences such as “It was not by the wolf_i that Jack_j suspected his_i,j mother_k had been eaten” ungrammatical unless “Jack” was coindexed with “mother,” and Poussin’s because it conflicts with subjacency.

³ Copies of the program are available from the author on request.

⁴ Velasquez (1985) has suggested that we view writing as a performance error. Another hypothesis might claim that there are mental rules systems governing language (spoken language) and writing, which, since it is not language, should not be examined by linguists.

	Discourse Gender in Hakka Creole—Keith Slater
	Perpetuation of Traditional Gender Roles by European Languages—Douglas S. Files
	Babel Vol I, No 1 Contents