五、关于词语指涉事物的“言语行为”说

六、余论：认识自己与人类未来

Noam Chomsky

A good place to begin is at the beginning, 2500 years ago, about as far back as detailed records go. That is when the Delphic Oracle issued a pronouncement defining our primary task: Know Thyself.

A two-word aphorism can be interpreted in many ways. Given what we know today, and what should be uppermost in our minds, a reasonable way to interpret the Oracle is in collective terms: as urgent advice to try to understand what kind of creatures we humans are.

And we are indeed strange beasts, at once the pride of evolutionary history and the scourge of the earth. It’s a challenge to find out how this could be. And an imperative to find out in time so that we can fend off the worst and aspire to the best.

Humans appeared very recently, some 2-300,000 years ago, a blink of an eye in evolutionary time. Not surprisingly, there is very limited diversity among us. The slight differences are what matter for life; the commonalities we take for granted. For understanding ourselves, the proper stance is the exact opposite: What is deeply important and most revealing is what is common to us, what distinguishes humans so radically from all other life on earth. Variation is superficial.

That much is true on pure intellectual grounds, but also in its import for human life. It is not news that we are at a unique moment in human history, a moment when decisions have to be made, quickly and decisively, as to whether human society will persist in any organized form. We are living at a moment of confluence of severe crises, existential ones. All are collective, with no boundaries. We will answer them together, or not at all, in which case the human experiment will come to an inglorious end. I’ll return at the end to a few words on that.

For good reasons, then, we should interpret the advice of the Oracle collectively: What Kind of Creatures are We? What are the species properties that are common to humans apart from severe pathology, and are without significant analogue elsewhere in the living world, even among our closest relatives the higher apes? These are the properties that paleoanthropologists call “the human capacity”.

When we inquire into this question, I think we find two striking properties that satisfy this austere criterion: language and thought—at least thought in any sense that we can grasp and can study. It is language and thought that enable humans to issue the proclamation of the Oracle, to reflect on its meaning, to seek to find answers to the questions it awakens in our minds. And the same is true of the most mundane experiences of ordinary human life, all common to humans and in important ways distinctive to the human species.

If inquiry reveals two distinct properties that virtually define the species, the next question is what relation holds between them. The simplest answer would be the relation of identity. Language is a system for generating thought, and thought is what is generated by language.

That in fact is the answer that was given at about the same time as the pronouncement of the Oracle: in classical India by a famous Sanskrit scholar, Bhartrhari, one of the founders of the great tradition of Indian grammar and philosophy. In his conception, “language is not the vehicle of meaning or the conveyor of thought” but rather its generative principle: “thought anchors language and language anchors thought… using language is thinking, and thought ‘vibrates’ through language.”

Similar ideas resonate through intellectual history. The 16th century Spanish physician-philosopher Juan Huarte set the stage for the first cognitive revolution that followed shortly after by emphasizing what he called “the generative quality” of human understanding, unique to humans: the capacity that we constantly exercise in normal life in using language to construct new thoughts and to interpret those of others. Huarte also identified a still higher form of intelligence that enables some to create work of true intellectual and aesthetic value—while the rest of us, lacking this talent, are at least able to enjoy and appreciate it, sometimes to carry it forward, another species property that probably derives from the language-thought complex.

Huarte also insisted upon the universality of language structure, on the brain as the material site for cognitive functions, and on the innateness of these functions. Along with the generative quality of human understanding, these ideas have been revived, without any awareness of the long-forgotten history, in what is called “the cognitive revolution” of the 1950s, in a sharp break from then-prevailing structuralist-behaviorist doctrines.

These ideas flourished during the 17th century scientific revolution, which laid the basis for modern science. A core part of the revolution was the willingness to be puzzled about simple things, ordinarily just taken for granted. Why do rocks fall the earth while steam rises? Why do we see a certain visual presentation as a triangle. Galileo and his contemporaries were not satisfied with the “occult ideas” that were invoked to account for what happens in the world, and wanted to focus attention on simple phenomena and find explanations for them. Much as children do with their incessant why-questions. That’s the stance that has always driven understanding forward.

Galileo held that nature is simple and it is the task of the scientist to demonstrate that. For Galileo, it was an ideal to guide research. In the centuries that followed it has been found to be true in so many domains that it has become a firm belief—what Albert Einstein called a “miracle creed”. In his words, “a miracle creed which has been borne out to an amazing extent by the development of science”.

Language did not escape the attention of the founders of modern science. Galileo and his contemporaries
expressed their awe and wonder about a truly remarkable fact: with a few symbols, each of us can construct infinitely many thoughts in our minds, and can convey to others with no access to our minds their most innermost workings. For Descartes, the normal creative use of language, employing this capacity, was a foundation for his postulating a second substance, res cogitans, relating language and thought, a substance unique to humans. Galileo himself
regarded the alphabet as the most spectacular of human inventions because it captured this marvel.

Galileo’s amazement was echoed centuries later by Gottlob Frege, the founder of modern logic and logical philosophy. Frege found “astonishing what language accomplishes. With a few syllables it expresses a countless number of thoughts, and even for a thought grasped for the first time by a human it provides a clothing in which it can be recognized by another to whom it is entirely new.”

The amazement of these great thinkers is entirely warranted. As is their insistence on puzzlement over what seems commonplace and self-evident. It rarely is, when we look closely.

The insight about language and thought poses what we may call “the Galilean challenge”: How is this achievement possible? The challenge seems to me to capture lucidly the major task faced by the inquiry into the nature of language and thought, in fact into the unique nature of the human species.

The Galilean challenge initiated the rich tradition of General and Rational grammar, general because it sought universal principles underlying human languages, rational because it sought to go beyond description to explanation. Throughout, language was generally regarded as closely bound to thought. The common understanding was captured simply in the phrase of the 19th century linguist William Dwight Whitney that language is audible thought—though we now recognize that sound does not have that unique status; other sensory modalities will do. The great humanist and founder of the modern research university Wilhelm von Humboldt, went further. He identified language with thought. He characterized language as “a generative activity” and he pondered the fact that somehow this activity “makes infinite use of finite means”—a basic feature of the Galilean challenge.

Humboldt’s formulation of the Galilean challenge brings to light a serious gap in the entire tradition. It failed to accommodate Aristotle’s crucial distinction between possession of knowledge and use of knowledge; in contemporary terms, between competence and performance. The tradition kept to use of knowledge, more specifically, use of knowledge in production of speech. Little was said about perception of language, a prime concern of modern psycholinguistics. Both perception and production access the knowledge that is internally possessed.

Appropriate means were not available to formulate clearly and pursue the fundamental task: to unearth the system of knowledge possessed, the internal language, I-language in modern technical usage.

This gap was overcome by the modern theory of computation created by Alan Turing and other great early 20th century mathematicians and adopted in the modern reshaping of the tradition. These tools made it possible to resurrect Aristotle’s forgotten distinction between possession and use of knowledge and to launch serious study of possession of knowledge for the first time. They also made it clear how a finite mechanism can have infinite output. That does not entirely resolve Humboldt’s puzzle about production, but it makes it possible to carve out the domain of feasible inquiry: the study of generation of the knowledge possessed, with finite means and infinite scope. Production of language, like other creative activity, is beyond reach of inquiry in any fundamental sense, as is even simple voluntary action.

The distinction between generation and production of language is fundamental, and commonly misunderstood. Generation lays out the terrain: here are the linguistically-formulated thoughts, the whole infinite range of them. Production is an activity that makes use of these thoughts, commonly in innovative and creative ways that are not understood. In Cartesian terms, our use of language is appropriate to circumstances but not caused by them. We are incited and inclined to speak in certain ways, but not compelled. And Huarte’s higher form of creativity is still more mysterious.

The new tools of the theory of computation made it possible for the first time to develop explanatory theories of a language that yield an infinite array of hierarchically structured expressions that constitute thought, and can be externalized in some sensory-motor medium, usually sound—what we can call the Basic Property of Language.

As soon as this enterprise was undertaken, problems arose. The reigning structuralist-behaviorist consensus had no real puzzles. Everything essential was known; procedures of analysis could be applied to any corpus of materials, yielding a structural description. Acquisition of language was understood to be just a matter of training and habit, as the leading American linguist, Leonard Bloomfield, formulated widely accepted doctrine.

As soon as the search for explanation began, it turned out that almost nothing was known. Virtually every sentence poses new puzzles. Furthermore, the elements that were postulated in the best explanatory theories could not possibly be found by a schedule of procedures, just as in every other branch of science. Search for the best theory is a creative activity. There are no algorithms for such endeavors.

Explanation in the language case has to proceed at two levels. A generative grammar of a language is a theory that seeks to account for its properties, the knowledge possessed by the language user. At a deeper level, the theory of the shared language faculty is concerned with the innate factors that make language acquisition possible—factors that distinguish humans from all other organisms. In modern terms, it’s called universal grammar, UG, adapting a traditional term to a new framework.

UG has goals that appear contradictory. It must meet at least three conditions:

We achieve a genuine explanation of some linguistic phenomenon only if it keeps to mechanisms that satisfy the joint conditions of learnability, evolvability, and universality, which appear to be at odds. The course of theoretical inquiry has been driven by the goal of reconciling these conflicting requirements.

These problems were evident at the outset of the generative enterprise. We now know that they are considerably more severe than was then envisioned.

With regard to evolvability, genetic studies have shown that humans began to separate not long after their appearance. There are no known differences in Language Faculty, which must have already been in place. Furthermore, there is no meaningful evidence of symbolic activity prior to emergence of Homo sapiens (modern humans). These facts suggest that language emerged pretty much along with modern humans. If so, we would expect that the basic structure of language should be quite simple, the result of some relatively small rewiring of the brain that took place once and has not changed in the brief period since. The apparent contradiction with learnability therefore becomes even sharper.

Research on language acquisition has extended the dilemma further. It has shown that a child of two or three years old has largely mastered basic properties of its language, including some remarkable ones. Furthermore, statistical study of the data actually available to a child shows that relevant evidence is very sparse. These discoveries seem to require that UG must be very rich in order to overcome the enormous gap between data available and knowledge attained, while the evolvability condition demands that UG be very restricted. And the problem of apparently endless variation lurks in the background.

In the past few years, some hope has emerged to resolve this tangle of dilemmas, within the so-called minimalist program. We can see how by looking at a fundamental and quite puzzling universal property of language: structure-dependence.

Consider the properties of this simple sentence, mastered by two-year-old children as experiment has shown:

Not “is in the room”.

That raises puzzles. To determine agreement, the child does not use the simplest computational rule, adjacency. Instead, the child reflexively relies on something it never hears: the structure its mind creates. The child then assigns plurality by virtue of the nature of this abstract structure.

The problem quickly extends. Let me take the sentences below as example:

To decide whether to use was or were, we ignore linear adjacency and attend only to the structure, the Noun Phrase subject. But we have to go beyond to find the central element of the Noun Phrase, friend/friends, the more remote Noun. That turns out to be a non-trivial computation, far more complex than linear adjacency, which is reflexively ignored.

Let’s take it a step further. Consider the sentence “The mechanic who fixed the car carefully packed his tools”. It is ambiguous, “fixed the car carefully” or “carefully packed his tools”. Now put the adverb “carefully” in the front: “Carefully, the mechanic who fixed the car packed his tools?” It is now unambiguous. It means that he carefully packed his tools. The adverb carefully seeks a verb, but it cannot use the simplest computation: pick the linearly closest verb. UG forces us to ignore that simple computation and to select the more remote verb—which happens to be the structurally closest verb as we see when we assign the structure in what our minds tell us is the correct way.

It goes without saying that none of this can possibly be learned.

There is supporting neurolinguistic evidence. Experimental work shows that if subjects are presented with invented languages modelled on actual ones, the language areas of the brain react normally; but if the invented languages use very simple rules involving linear order, there is diffuse activity in the brain, indicating that it is being treated as a puzzle, not activating the language areas.

We therefore have a serious puzzle. The infant ignores 100% of what it hears—linear order—and attends only to what it never hears, abstract structures that its mind constructs. Furthermore, this is true for all constructions in all languages.

The only plausible answer is that linear order is simply not available to the child who is acquiring a system that constructs thoughts. Why then does speech require linearization? The reason is obvious. The articulatory system cannot produce structures, so the externalization process must impose linear order on an internal system of generation of thought, which is unordered. Sign language is less strictly linear because of wider options available in visual space.

The sensory-motor systems used for externalization have nothing to do with language; they were in place long before language emerged, and have not changed since. True language is the internal system that generates thought, the I-language—internal to an individual, coded as a recursive generative procedure in the brain.

We now face the question why the infant reflexively ignores 100% of what it hears—linear order—and attends only to what it never hears—abstract structures constructed by the mind.

We’ve recently discovered that there is a straightforward answer. Satisfaction of the Basic Property requires a computational procedure. The simplest procedure is binary set-formation, called “Merge” in recent work. It is easy to show that if recursive generation is based on Merge, then structure-dependence is imposed and there is no linear order. In this respect, language conforms to the miracle creed, and seeks the simplest solution.

There are many consequences. One consequence is that certain thoughts become inexpressible in speech, even very simple ones. To illustrate, let’s shift from conjunction to disjunction. Take the sentences “Either John or Mary is in the room” and “Either the girls or the boys are in the room”. How about “Either the girls or John [is or are] in the room”? Neither is possible. The thought is clear, and if number agreement were based on adjacency it could be easily expressed. But with structure-dependence, the thought is inexpressible in the simplest way, because of the internal conflict within the disjunction.

This is one of many examples of conflict between optimal design and utility for communication. There are many others that are far more serious than this one. They are always resolved the same way: communicative efficiency is always sacrificed. To put it metaphorically, when Mother Nature was constructing language, she was concerned with optimal design, not how the system might be used. If it turns out to be dysfunctional, so be it.

It’s worth noting that that’s how evolution works quite generally. We can distinguish three stages in normal evolutionary change. The first stage is innovative. Some disruption takes place, maybe mutation or gene transfer or a bacterium accidentally swallowing some microorganism, leading to eukaryotic cells, the basis for complex life. The disruption introduces something new. Then comes the second stage, reconstruction: nature re-designs the new system in the simplest way, satisfying natural law. These are matters discussed in depth in classic work of D’Arcy Thompson on laws of growth and by Alan Turing on creation of patterns by physical law. It is the miracle creed once again. The final, third stage of evolution is winnowing: natural selection reduces the variety to the better adapted. The simplest solution produced at the second stage might turn out to be dysfunctional, but nature doesn’t care. In reconstruction, nature seeks optimal design. It has no way to consider possible functions.

Language seems to fit the normal pattern. We can envision an evolutionary scenario that looks something like this. Some small rewiring of the brain yielded the general property of recursive generation. Nature then took the usual course of seeking the simplest such operation, relying on natural law (in this case, principles of computational efficiency, since we are dealing with a computational system). The result is a system based on Merge, which satisfies the Basic Property. The deepest and most surprising property of language, structure-dependence follows at once, with its ramified consequences. UG encodes the outcome. The infant therefore reflexively ignores everything it hears and attends only to the abstract structures its mind creates, fleshed out as thoughts.

In these terms, we can envision a resolution of the conundrum facing UG: the diversity condition (iii) can be put to the side in the study of the core system of generation of thought. The variety of languages might be localized largely in peripheral aspects of lexicon and in externalization; perhaps completely localized there, we might someday learn. Research seems to me to be tending in that direction, and it would be a natural outcome. The poverty of stimulus (POS) conditions on the core system are severe, sometimes clearly insurmountable, as in the case of structure-dependence. Furthermore, the Sensory-motor (SM)/I-language relation poses a cognitive problem that can be solved in many ways. It would not be surprising to find that it is the locus of the complexity and mutability of language along with their variety of languages—all superficial if these ideas are on the right track.

If so, at least for I-language, the conundrum is largely reduced to satisfying the evolvability condition (ii). That problem will be overcome to the extent that the structures of I-language are generated by the simplest operations. The Strong Minimalist Thesis (SMT) sets this outcome as a prime goal of the theory of language. The assumption again is that language conforms to the miracle creed along with the rest of nature in its core property of construction of thought.

These observations reinforce the traditional view that language is essentially a system of thought. The modern doctrine that language is basically a system of communication that may have evolved from animal communication that seems quite untenable.

Returning to the evolutionary scenario, the third stage, the winnowing stage apparently has never been reached, though it seems to be no diversity. Possibly that’s true, because the time has been so short, more interestingly, because the system is so tightly integrated that it’s either all or none. That seems to be increasingly what we find when we investigate more complex cases.

Instead of going on to further cases of genuine explanation, let’s turn to how thoughts are fleshed out.

I-language is a computational system. Any such system has primitive elements—atoms of computation—and computational procedures to form new elements.

For language, the primitives are the minimum “meaning-bearing” elements. In an analytic language like English, these are often words, so you have major books with titles like Words and Objects (W. V. Quine) or Words and Things (Roger Brown). For language generally, word is not the right concept, even for English if we look closely, but let’s keep to that.

The titles of the books reflect the standard view that words are linked to mind-external objects by the relation of reference: words refer to things in the world. For the sciences, establishing such a relation is a norm. When scientists speak of Uranium or genes, they hope there are such things. It also seems that animal communication systems observe these conditions, with a 1-1 relation between the signals, say a monkey call, and some physically identifiable event. But there is very strong evidence that human language does not observe the referential doctrine. We of course do use words to refer. That is an action, what the British philosopher John Austin called a speech act. When we look closely at the meanings of words, it is hard to sustain the idea that there is a word-object relation.

Take a standard example in the philosophical literature, a supposedly, logically proper name, like London. I can perfectly well say that I’m going to visit London again after it was destroyed by fire and rebuilt 50 miles up the Thames with different materials and design. No one believes that there is a real-world entity with such properties as these. The word London is not a proper name, it’s a city name, and our mental concept city has rich and complex properties, carried over when we use the word London in the speech act of referring.

The observation generalizes to the whole vocabulary. The basic point was recognized by Aristotle. Aristotle discussed the word house. He observes that house is an amalgam of matter and form. The material part is what a physicist could identify: bricks, timber, and so on. The form is a mental construction: the design, the intention of the architect, the characteristic use, and so on. Something could look exactly like where I live but not be a house, could be a library, could be a garage, could be a paperweight for a giant. These are not properties determined, detectable by physical examination.

Even earlier, a similar point was made by the Greek philosopher Heraclitus when he asked how we can cross the same river twice, though its material component has changed. It’s a deep question. A little thought shows that under radical material changes it remains the same river, while under slight changes it might not be a river at all, could be a canal, or with a hardened surface, could even be a highway. That turns out to be true of even the simplest words that we use to refer.

These features of language and thought raise new conundrums. All of this must be known without learning, and correspondingly seems to be shared cross-linguistically in fundamental respects. Studies of language and language acquisition have shown that these items are acquired on very few presentations, implying that they are based on rich innate structure. Learning is apparently a matter of settling relatively superficial properties. There are no significant analogues to such concepts in animal symbolic systems. Of course, we have no record of their development.

For these reasons, the origin of human concepts has seemed a hopeless mystery, buried deep in the pre-history of homo sapiens. But on reflection, that seems most unlikely. The rich and intricate concepts that are the atoms of computation would have no function outside of a system of generation of thought that yields complex expressions that can enter into reasoning, reflection, and other mental acts. If they had emerged before language, they would have been useless waste and would have been discarded. So, it is reasonable to suppose that when the capacity emerged for recursive generation of thought, it appropriated a lexicon of very elementary concepts available to proto-humans; and only later, as the combinatorial possibilities of generating language and thought became available, concepts of the distinctive human character appeared—how, we can only guess, but at least it might be possible to explore the question.

Here again the role of the Strong Minimalist Thesis (SMT) might appear. When some innovation appears, nature seeks the simplest way to accommodate it. The simpler the factors that do so, the more refined will be the system that is produced. With many factors intervening the result is likely to be chaotic. It is not too much of a stretch of the imagination to think that the simplicity and discipline of the Strong Minimalist Thesis had a significant role in the development of these unique conceptual resources of human thought. As more is being learned about the rich, inventive, creative lives and complex social orders of the tens of thousands of years of human history that are now becoming accessible to inquiry, new and helpful insights into these questions might be coming into view. It will be exciting cross-world.

Following these paths, we may hope to gain real insight into the nature of language and thought, fundamental features of the human capacity. Fundamental, but of course not all inclusive. When we look around us, we see other striking features of this strange creature that recently appeared on earth, those features I alluded to at the outset, having to do with the existential crises that we face and must overcome soon.

The issue arose with dramatic clarity 77 years ago, on Aug. 6th, 1945. On that day, we learned that human intelligence had reached the level where it had found ways to destroy us all—actually not quite yet, but it was clear that science would soon reach that point. And it did, in 1952, when the US exploded a thermonuclear weapon, soon followed by the Soviet Union.

That raised a profound question about human nature: can our moral capacity rise to the level where it will control our capacity to destroy? The record of the years that follow is not encouraging, not just with regard to nuclear weapons. It was not known in 1945, but the world was entering into a new geological epoch, the Anthropocene, an epoch in which human activity has a grave impact on the environment, so serious that it imperils the survival of organized human life on earth, not to speak of the vast number of species that we are wantonly destroying.

The World Geological Organization has dated the onset of the Anthropocene to the period after World War II, when the destruction sharply escalated—By now, it has reached the level where irreversible tipping points are in sight.

The question has only become more stark over the years. There is a gap between the human capacity to destroy and human moral capacity to control their impulse. Can that gap be overcome? We will soon find out, and if it’s the wrong answer, nothing else is gonna matter.