Charles Hockett Critiques Noam Chomsky

Charles Hockett, The State of the Art, 1968 (The Hague: Mouton).

Hockett critically reviews the view of Noam Chomsky. His general approach: "those views are largely in error, but they are two powerful to be shrugged aside. It is necessary to meet Chomsky on his own ground. When we do this, we discover that, even if he is wrong, his particular pattern of error tells us some things about language were were formerly unknown" (from the preface).

Chomsky's central fallacy: "gazing on stability and seeing well-definition."

Notes about what this document is: In the late 1980s or early 1990s I read Hockett's book and took these notes for my own uses, from a library book because it was impossible, in those days before, to buy the book. This ideas are more relevant than ever, now that even formal linguists may be asking if statistical pattern-recognition techniques, and documented human regularity extraction abilities, could be relevant to questions of how children learn language. An example of this tendency is MIT's October 19 2007 workshop, "Where Does Syntax Come From? Have We All Been Wrong?" -- Notes by Catherine Caldwell-Harris, October 21, 2006.

This is a condensation of key ideas, mostly in the form of direct excerpts, with some paraphrasing, follows.


Chapter 1: Surveys development of linguistic theory in the US from 1900-1950.

Chapter 3: Discusses the background framework for Chomsky's central assumption, whether Language is, as Chomsky claims, a well-defined formal system.

Chapter 4: Even accepting the framework discussed in Chapter 3, there are alternatives to the view that language is a well-defined system
Chapter 5: Critique of the Chomskyan view.
Chapter 6: What Hockett believes is the correct approach.

Page numbers should help readers find the original paragraphs, as these should be consulted for quotation purposes.

(From middle of Chapter 3) Is language a well-defined formal system?

Hockett summarizes the Chomskyan view in 19 statements, labeled c1-c19. Repeated here are the 2 he refers to most frequently.

C10. The grammar of a language is a finite system that characterizes an infinite set of (well-formed) sentences. More specifically, the grammar of a language is a well-defined system.

C14. The innate grammar-producing system is a well-defined system in the sense of C10, just as is the grammar of any single language.

p. 47 Hockett gives examples of what can and can not be characterized as a formal system.

Chess can be characterized as a formal system.

Why? A given distribution of pieces on a board is a state. The rules specify a fixed initial state. They specify the means of moves and captures, and exactly what states can immediately follow any given state (the permissible state-transitions). They specify what states are terminal, and assign all states to three pairwise disjunct classes: white wins, black wins, and draw. A game of chess is a finite sequence of states of which first is the initial state, the last is a terminal state, an each state-transition is permissible. Two games are formally identical if they are identical state by state.

Baseball can be characterized as a formal system.

Why? Because the rules and umpires sharply quantize balls, strikes, fouls, outs, etc. Ignoring points of audience interest and the demeanor of players, we can define formal identity between actual games, and enumerate all possible distinct games. The set of all possible baseball scores (earned by one team in one game) is also known: the set of non-negative integers.

American football can not be characterized as a formal system. (p. 47) Why not?

The set of all possible scores that can be earned by a single team in a single completed game is undefined. A safety wins 2 points, a field goal 3, a touchdown 6, a conversion after a touchdown 1 or 2. Any score then must be of the form a + 2b + 3 + 6d, where a, b, c, and d are nonnegative integers and a <= d.

But scoring requires time and play to be confined, at most, to slightly more than 60 minutes of time. A score of 1 million is impossible. The highest score on record is 227. Could it be higher? We have no way of saying what the maximum score could be. This set of all possible scores is thus ill-defined.

American table manners (p. 51) can not be characterized as a formal system. Although we can codify "rules" of proper manners, these don't meet the criteria of a formal system, and must be considered guidelines or heuristics.

p. 52. Hockett asserts that no physical system is well-defined.

Physical systems have moving parts. Moving parts mean random wear and tear ("thermodynamic indeterminacy"), the effects of which are predictable only in a gross statistical way. If the system is not isolated, then unforeseeable influences from outside will impinge, whose effects are again only statistically predictable.

p. 54. Hockett provides an interesting comparison of a physical system (not well defined) and the well-defined notational system which is used to describe it.

Consider the relationship between hydrocarbon molecules of the methane series, and the structural formulae which describe the hydrocarbons. To construct a formula of this set, take N c's, 2n+2H's, and 3n+1 short line segments, which n is any positive integer. Arrange the C's in a row and insert a line segment after each but the last. Attaching H at each end, one one H above an done H below each C. Thus, for n=1, 2, 3 we get the formulae for methane, ethane and propane.

Nothing (but time and motivation) prevents us from writing down the formula for the case of n=1,000,000 or more. The set of formulae is well-defined, and infinite.

Not so the hydrocarbons themselves. As n increases, the chain twists in space more and more and stability decreases. The probability that a methane-series hydrocarbon molecule with large N will actually form decreases as n increases, as does the life-expectancy of such a molecule if it does form. There is no well-defined "longest possible" methane-series hydrocarbon molecule; yet the number of chemically possible distinct hydrocarbon molecules is not infinite. This is a fact left out of account in the chemist's system of notation, just as that notation requires only two dimensions as over against the three occupied by the molecules, and just as it represents, say, a carbon atom by the completely nonpictorial symbol C.

This pair of examples is instructive in that it shows us a useful, though only approximate, matching between a well-defined system and an ill-defined system. The organic chemist manipulates hydrocarbons in test tubes; he manipulates formulae on paper or blackboard. The second kind of manipulation enables him to work out hypotheses, which can then be tested in the laboratory with the actual substances. Only the latter qualify as findings. If a hypothesis and a finding conflict, it is the hypothesis that is modified or discarded. If the chemist's terminology and notation yield too many hypotheses falsified by experiment, they are revised. The fact that the set of all formulae for hydrocarbons of the methane series is a well-defined system is a discrepancy between notational system and reality; but it is a harmless discrepancy that has probably never mislead any working chemist.

What Hockett implies here is that working linguists have been mislead by their notational system. They have come to believe that the notational system is not just a useful descriptive device, but is isomorphic (the same thing) to the knowledge structures that mediate human language.

Chapter 4

p. 56 Consider the relation between notational system and subject matter in different scholarly disciplines.

In chemistry and physics, no one confuses these.

Consider logic and pure mathematics. Here there is also no confusion of terminology and subject matter, but for a different reason. In these disciplines, the terminology/notational system *is* the subject matter. (But note exceptions below)

Is linguistics an empirical science like chemistry, or a formal discipline like logic and mathematics?

Hockett's firm conviction is that linguistics is an empirical science. He notes that the alternative was/is attractive to some temperaments. Chomskyan linguistics is only rendered plausible if one accepts that language is a well-defined (i.e., formal) system.

Let's not take this proposal on faith, but instead, seriously entertain it is as hypothesis. Are the consequences in accord with experiential and experimental fact? p. 57.

Some reasons to believe that language might be a well-defined system:

1. The only existing well-defined systems are the inventions of human intelligence. One proposal about such systems is that they arise through certain uses of language and its immediate derivatives such as writing. This proposal requires that language be a well-defined system. Let us call this hypothesis the "Law of Conservation of Well-Definition" i.e., "only a well-defined system can give rise to a well-defined system."

If true, then a language in turn must come from a well-defined source. But neither genes nor cultural transmission are well-defined.

Rejecting the Law of Conservation of Well-definition, let's look for ways in which well-definition can arise from ill-definition. Where is the step from ill-definition to well-definition? Three possibilities: (described on p. 59.)

1. Even though physical genetic transmission and cultural transmission are ill-defined, they still give rise to a well-defined grammar producing system in the child (so C10 and C14 both hold).

2. The grammar-producing system of the child is ill-defined, but it can yield well-defined grammars (so only C10 holds).

3. Languages themselves are ill-defined, and well-definition *only* appears through special uses of language that give rise to things like mathematics.

These 3 categories collapse to two: either language is well-defined or it isn't. Chomsky has chosen the former; Hockett chooses the latter.

Chapter 5

Hockett debates Chomsky's 19 points

Hockett argued against 17 of the 19 points he developed to concisely summarize Chomsky's view (as stated in 1965). (Chomsky did read through the 19 points and basically agreed that those were in fact his points, although see qualifications on p. 38.)

C1. The vast majority of sentences encountered in everyday life are encountered only once; that is, most sentences are novel. (Hockett does not dispute this one although analyses of text indicate that most sentences are either exact copies of well-known sentences or are minor variations of commonly heard sentences.)

C2. Any user has access, in principle, to an infinite set of sentences. In practice, all but a finite subset of these are not usable to the listeners, either because they require too much production time, are too complex, are false or are meaningless.

Hockett view this to be absurd. Just as a million is not a possible football score, a sentence containing a million repeats of "one and one and" is not a possible sentence in English.

Recall the hydrocarbon series example. As one attempts to stick more legal language pieces onto the sentence, one encounters flexible constraints. These constraints are *part of the language*, just as the time limits of a football game are part of football. Moreover, it seems that *all* constraints in language are of this more or less of this rubbery sort.

Hockett envisions a reader who might ask: Even if languages are not well-defined, isn't it useful to create a formal approximation to it, as in the notational system used to describe hydrocarbons?

p. 62 Hockett replies: This depends on your goals. An approximation always requires leaving something out of account. The approximations we obtain when we characterize natural language as a formal system leave out just the most important properties of human language, in particular, they leave out the source of language's "openness" (what allow novel sentences to be generated and understood).

p. 63. Before Chomsky, linguists spoke of skills and habits, rather than of knowledge. Why is this?

The English verb "know" means two things: *Know how to* (know how to walk, know how to speak), and "have knowledge of" (have knowledge of facts, rules or principles in some domain).

The average man has little knowledge of the muscular mechanisms of walking, yet knows how to walk. Why claim that native speakers "have knowledge of" their language, when they have no knowledge of the rules or principles that underlie language habits?

Chomsky inveighs against using the terminology of "habit" on the grounds that habits, skills, routines have no established sense in which it can explain language competence or the acquisition of language. But today, do cognitive scientists have the means to explain acquisition in terms of habits, skills, routines?

Competence-performance distinction

p. 64 Chomsky notes that his C-P distinction is similar to the classical langue-parole (language-speech) contrast.

The classical view is that speech is what speakers actually produce. Language is the system of skills which produces speech.

Chomsky posited that speech and language are independent entities, each of which is worthy of study in its own right. Hockett views this to be preposterous.

When we observe a specific historic event, be it a speech act or otherwise, we can talk about it in two ways. We can be specific, or we can try to generalize. But there is only one "object of study:" specific acts of speech, as historic events, in their behavioral settings, observable in part overtly and in different part introspectively.

Accurate reports of observations are not theories: their sole importance for theory is that they enable the theorist to examine the evidence at leisure. The linguist seeks theories, which are generalizations from observations, and are about speech. They yield predictions, and are corrected by subsequent observations.

Empirically, we might be led by our observations of speech to propose that the underlying system, the set of habits we call the language, is well defined. But Chomsky was not led to this conclusion empirically, nor has anyone proposed this as an empirical hypothesis.

The assumption of well-definition can be retained in the face of the evidence if one posits an obscure sort of "underlying" system that by definition meets the requirements of the assumption, and then explains (or explains away) the vagaries of actual speech as due to the participation of other factors. But this step moves the underlying system completely out of the reach of the methods of empirical science. The notion thereby ceases to be a hypothesis.

What about Chomsky's "Ideal speaker-listener"?

p. 66 There is nothing wrong with employing idealizations in a theory, but we must remember what an idealization is. It is not what we are analyzing, not part of our subject-matter; rather it is part of the terminological apparatus with which we analyze and discuss real objects and systems.

Once we abandon the notion that a language is well-defined, the idealization of a speaker becomes useless. Much more useful would be to refer to the average or typical user of a language, who has, in full measure, all the `faults' of which Chomsky divests his imaginary ideal.

The relevance of probabilistic considerations

Hockett (p. 39) puts forward Chomsky's claims as follows:

Probabilistic considerations pertain to performance, not to competence. Surely the user of a language has knowledge of probabilities, but this knowledge constitutes a mental reality distinct from the grammar of the language. Knowledge of probabilities may influence performance, but there is no reason to believe it has anything to do with the organization of grammar.

Hockett's response: (p. 67)

Linguists have rarely used formal statistics, but in generalizing from observed speech they usual distinguish between varying degrees of productivity of patterns -- in Sapir's phrase, differing configurational pressures.

How is this relevant for understanding how people can say new things?

Consider the situation in which various partly incompatible patterns are all apt. Their interplay and the resolution of incompatibilities can lead to a sentence that has not been said before.

Chomsky's idea of "explanation" p. 67

Chomsky has a tendency to `explain' complexities by positing, ad hoc, more or less independent pieces of mental furniture. As in nineteenth century German `Greistewissenschaft' this actually explains nothing at all, but merely forestalls inquiry; though in fairness we must admit that in Chomsky's case the practice may simply be a device for filling various difficult matters for later consideration.

Chomsky's separation of grammar and semantics

p. 68 Here Chomsky prolongs an error of the 1940s.

Hockett gives the standard examples against this separation, including some interesting historical examples.

Interesting Bloomfield quite on page 71.

Chomsky's notion of meaningfulness (p 72)

Looking for his guidelines in logic and philosophy, Chomsky has imposed on language a definition of `meaningfulness' that is entirely out of place.

In logic, *two plus two is yellow* is meaningless, unless yellow has been assigned the the value 4.

What has all of this got to do with ordinary utterances? Practically nothing. In everyday use of language, people find it risky to decide whether an utterance is true before they find out what it means.

Hockett's own views

One of Hockett's main reasons for being suspicious of Chomsky's views was that they conflicted with what had been discovered about linguistic change.

p. 82 This is not mere ancestor-worship on my part. We all reject alchemy and astrology, despite the fact that hundreds of brilliant scholars devoted their lives to those subjects. At the same time, the fact that Tycho Brahe and Sir Isaac Newton retained astrological beliefs does not prevent us from accepting their real discoveries. Science is cumulative; this does not mean that we accept the findings of predecessors uncritically. My conviction that analogy, borrowing and sound change are the major mechanisms of linguistic change is not predicated on a blind faith in Brugmann, Leskien and company. Rather, my respect for them is based on the fact that, by paying honest attention to the evidence, even it forced them to set their own personal predilections aside. They found that the evidence overwhelmingly supported the hypothesis -- as, indeed, it still does.

p. 83 A language is a kind of system in which every actual utterance, whether spoken aloud or merely thought to oneself, at one and the same time by and large conforms to (or manifests) the system, and changes the system.

If a man takes a snapshot of a horse race, he must not be surprised that the horses in the picture are not moving. We get a stable picture of the language because that is what we have sought. Even if our work is very expertly done, we must not promote our more or less standardized by-and-large characterization of the language to the status of a monolithic ideal, nor infer that because we can achieve a fixed characterization some such monolithic ideal exists, in the lap of God or in the brain of each individual speaker.

p. 85. "sand lot chess" -- inhabits of an isolated village play chess their whole life, only the rules are never written down and all learning is via observation and participation. Furthermore, the "rules" can be changed -- they are whatever you can get your partner to accept.

"To our way of thinking, sand lot chess is not nearly so desirable a game as real chess. It is not very much like a language. But it is much more like a language than is real chess."

to Catherine Caldwell-Harris' home page