Empirical Constraints for Universal Grammar

If you are really interested in this kind of stuff, don't use my paper for citations ; rather, use the references I cited as a jumping-block, since they are much more interesting and profound than this work. I also ask that you respect the copyright notice and provide a reference to this page when citing material. Thanks.

Empirical Constraints for Universal Grammar

by Leif Morgan Johnson
May 2001
for PSY 499
supervised by Dr. James Kalat
North Carolina State University

Abstract

After Chomsky's initial formalization of the concept of a Universal Grammar in the 1960's, the idea received both harsh criticism and enthusiastic support. Chomsky proposed a methodology for studying languages, but his ideas had little empirical support. Since then—and thanks in part to his departure from behaviorist methodologies—cognitive and developmental psychologists have collected large amounts of data regarding language development and use in humans worldwide. These data include behaviorist and cognitive studies of adults and children, studies of patients with certain brain disorders or damage, and new technologically feasible studies using PET scans, CAT scans, and MRI's. When taken together with evidence from neuroscientists about the development and evolution of the human brain, these empirical data seem to constrain the possible nature of a Universal Grammar.

This paper presents evidence from a few major authors in the field of linguistics and language development. It brings together some ideas on languages from different disciplines, notably linguistics, psychology, and anthropology. The unity of these ideas sheds light on some possible explanations—in evolutionary and neurological terms—of the existence of some sort of a Universal Grammar, however far it might eventually be from Chomsky's initial formulation.

Introduction

Language use is a characteristic of humans that seems to be unique to our species. Not even other primates, with whom we share so many traits, possess the ability to master language use on a level comparable to that of human children, much less adults. An orangutan or chimpanzee cannot verbally boast to us or to its brethren about its recent exploits while foraging; dogs are limited to barks of warning, greeting, excitement, or disappointment; and the concept of a lab rat complaining to its cage-mate about the cramped living conditions is more humorous than anything else.

Given the unique nature of human language, it may seem surprising that a scientific study of language did not really begin until the nineteenth and twentieth centuries. Until about 1900, study of language was relegated to historians (on one hand) and philosophers (on the other). Philosophers made some significant advances in the first 60 years of the twentieth century, however, giving momentum to the relatively new science of linguistics.

This paper begins with a short description of the contributions to linguistics from Ferdinand de Saussure (around 1910) and from Noam Chomsky (around 1960). Then, with a concept of some of the characteristics of a language, the paper goes on to present some evidence from more recent studies in cognitive science, developmental psychology, and anthropology regarding language development and use from ancient times to the present. Such evidence and contributions from a variety of different fields provide a compelling perspective for questions about the grammar and function of human (natural) languages.

Finally, the paper concludes with a summary of the ideas presented and how they affect or constrain the possibilities of a Universal Grammar, both theoretically and in practice.

Early linguistic developments

Saussure's linguistics

Until about 1900, the study of languages was dominated by what Ferdinand de Saussure calls diachronic linguistics, also referred to as historical linguistics, evolutionary linguistics, or loosely as etymology [Saussure 86]. Certainly, some philosophers proposed theories of grammar, linguistic meaning, and societal function, but very little attention was paid to what we now call a formal grammar for human (natural) languages. Saussure's ideas about the meaning of a linguistic sign and systems of symbols revolutionized thought about linguistics.

In particular, Saussure was one of the first to propose a conceptual division between what he calls the linguistic signal, the linguistic signification, and the linguistic sign; that is, the theoretical pattern of sounds that represents a concept, the concept or idea itself, and the two considered as a whole. This is not to say, however, that they are separate entities: Saussure explains that his terminology has the advantage of indicating the distinction which separates each from the other [i.e., the signal from the signification] and both from the whole [i.e., the sign] of which they are part. (p. 67)

Saussure also proposed the important distinction that linguistic symbols have value only with respect to other symbols in a linguistic system:

If we say that these [societal] values correspond to certain concepts, it must be understood that the concepts in question are purely differential. That is to say they are concepts defined not positively, in terms of their content, but negatively by contrast with other items in the same system. (p. 115)

So, in general, a given word can refer to some idea, but the true sense of a word (its societal or linguistic value) is only determined by its relationship to other words in a given language.¹

Chomsky's Universal Grammar

In the late 1950's and early 1960's, Noam Chomsky developed a comprehensive new theory of linguistics based on grammatical analysis:

Let us define Universal Grammar (UG) as the system of principles, conditions, and rules that are elements or properties of all human languages not merely by accident but by necessity—of course, I mean biological, not logical, necessity. Thus UG can be taken as expressing the essence of human language. UG will be invariant among humans. UG will specify what language learning must achieve, if it takes place successfully. ... Each human language will conform to UG; ... UG will specify properties of sound, meaning, and structural organization. (pp. 29–30)

Thus, Universal Grammar has the following characteristics:

1. UG is a system of rules, principles, and conditions for generating and understanding (learning) language.

2. UG covers all aspects of language: sound, meaning, and grammar.

3. UG is a constant entity that is the same for all humans.

4. UG biologically precedes language use in any human. This implies that human children possess UG before or at the moment when they begin to speak.

The problem that this paper addresses specifically is that these ideas regarding the existence and nature of Universal Grammar in humans now have a great deal of constraint due to the collected empirical evidence on the subject. Upon reaching the conclusion of this paper, the reader will learn that Chomsky's Universal Grammar is reduced to a small remnant of its original conception, although some of his ideas do indeed stay alive and probably will for quite some time. In particular, item (1) remains but is reduced to a very abstract portion of its original meaning, as long as items (3) and (4) are taken to be true.

Storing language

Although Chomsky, like Saussure, invented a system of thought that has altered the way people study languages, these ideas were and are largely theoretical. Data must be collected to support or refute any testable theory. (Pinker humorously classifies Chomsky as a paper-and-pencil theoretician who wouldn't know Jabba the Hut from Cookie Monster ([Pinker 95], p. 52) as he gives an example showing that it is necessary to have both theoreticians and experimenters.) This section presents some of the ideas that motivated the development of the concept of UG, particularly the conundrum of storing infinite language capabilities in a finite volume (the brain), and the interesting aspects of language that reveal themselves in bilingual and multilingual individuals.

Infinite amounts of language in a finite space

One of the major influences that drove Chomsky's development of the Universal Grammar was a fundamental philosophical problem that has plagued thinkers for ages: how can humans, who clearly have a finite capacity in other areas (e.g., physical movements such as soccer skills), generate an infinite variety of words and sentences? Even children who are four years old display a capacity for language generation that far exceeds even the capacities of two- and three-year-olds ([Pinker 95], pp. 269–271). But the most surprising part of language development is that children are capable of producing a vast array of words that they have never heard before. In fact, children are capable of generating seemingly complete grammars based solely on some limited sound patterns coming from parents, friends, and other humans nearby. Chomsky refers to this as the argument from the poverty of the input (cited in [Pinker 95], p. 42), and it is this under-determination of data that is the most perplexing. Surely children and adults do not generate words at random from the ether; the brain must be involved somehow.

As an illustration, Pinker describes a study in which children are presented with a picture of an unfamiliar object and told that it is a wug. The children are then shown a picture of two of these objects and asked what they are. Most children respond with the morphologically correct wugs even though the psychologist can be reasonably sure that no child has ever heard the exact word wugs previously. ([Pinker 95], p. 127)

For a neurological psychologist, this ability is particularly perplexing from a memory perspective. It is simply amazing that humans can generate a literally infinite (well, limited by the amount of breath one can store in the lungs) string of sounds to create a sentence that is grammatically correct—and that usually conveys some meaning to a listener. For example, by taking the simple sentence, The dog eats a cookie, one can easily extend the sentence to generate a new thought: The dog eats a cookie under the table. This process can continue ad infinitum by adding more prepositional phrases, adjectives, adverbs, and so on, as described in detail below.

Certainly, the ability to use language is not simply stored in memory as a large association of sound patterns paired with motor programs. Even with ten billion neurons in a human being's central nervous system, the number of connections among neurons is still finite. Since memory most likely occurs through the mechanism of long-term potentiation (LTP) at or around synaptic gaps, there are a finite number of data elements that could possibly be stored by such a system. A large number of these connections must be devoted to storing motor routines (particularly in the cerebellum), so the number of connections available for storing language must be very small compared to the linguistic ability of any fluent human. [Kalat 98]

A good solution to this problem is to require sentences in a language to have a similar structure. Chomsky's used this idea when he proposed that UG is a system of sentence-generation rules. These rules govern the possible syntactic structure of the linguistic expression of thoughts and take much less storage space than would be required to store all possible sentences in a language. (Indeed, any finite storage space is less than the infinite space required to store all of a natural language!) In English, for example, a simple, singular, present tense sentence might be structured as an optional determiner (article) followed by a noun and then a verb, which could be stored as one rule. Then, by storing this one rule and a look-up table of nouns and verbs, one could generate sentences like, Mom reads, The dog eats, and The hyena laughs.

This process can and does get even more complicated. In English, for example, a basic present tense sentence is composed of a noun phrase (NP) and a verb phrase (VP). A noun phrase, in turn, can be composed of an optional determiner (an article), an optional string of adjectives (A's), and a noun (N) followed by any number of optional prepositional phrases (PP's). A verb phrase is composed of a verb (V) followed by an optional noun phrase. All of this is shown as a Toy Grammar in Table 1. [Pinker 95]

Table 1: A Toy Grammar for English sentences. Optional elements are indicated in parentheses, and terminal elements are represented with lower case words. Adapted from [Pinker 95].

S	-->	(NP) (VP)
NP	-->	(determiner) N (PP)
VP	-->	verb (NP)
PP	-->	preposition NP
N	-->	(A) noun
A	-->	(A) adjective

The major benefit of using such a system of rules for language generation is the possibility of generating infinitely many sentences with a finite number of rules. For example, starting with S, we substitute NP VP, which could then expand to NP eats NP, which in turn could become The N eats a N PP, becoming The dog eats a A cookie under N, and so on. In this toy grammar, infinite sentences are possible by stringing together any number of adjectives, or by adding any number of prepositional phrases.

This process of stringing together arbitrary numbers of prepositional phrases is abstractly referred to as recursion. Recursive processes are powerful approaches in computer science, logic, mathematics, and other disciplines. Recursion reduces the amount of space necessary to specify the solution to a problem of arbitrary size. The key to recursive processes is that the structure of the problem must resemble itself; that is, for example, the NP that is part of a PP must be the same structure as the NP that makes up the subject of the sentence as a whole, even though they serve different functions in the sentence.

Pinker explains that there are similar rules for morphology (word generation). In his wug experiment, for example, Pinker's conclusion is that children of a certain age have learned the morphological rule [to] form the plural of a noun, add the suffix –s. ([Pinker 95], p. 127) Thus, children who have learned only this rule for plurals think they can pluralize any noun by adding –s, which works wonderfully for the case of pluralizing wug to wugs. This is also a compelling explanation when looking at children's errors in pluralizing some words, such as changing house to houses (correctly) but then mouse to mouses (incorrectly). In this case, the brain must somehow keep track of irregular words through some sort of look-up table similar to that used to store nouns and adjectives. [Pinker 95]

Universal Grammar, then, is often thought of in terms of a generational grammar like that described in Table 1. The question that arises then is whether all languages use the same grammar. Clearly not! If that were the case, learning a second language would be trivial, perhaps simply a matter of learning a new set of vocabulary words to plug in for nouns, adjectives, and verbs. Instead of uniformity, the variety of language structures worldwide is astounding. While English uses subject-verb-object (SVO) word order, other languages use different orders like SOV (e.g., Japanese) or VSO (e.g., Gaelic). Some languages (e.g., German and Russian) use declensions to indicate the function of a word in a sentence, while others (e.g., English) use a more rigid word order in sentences to preserve functionality. [Pinker 95]

Different languages in one brain

One of the major questions for which Universal Grammar fails is that of how it is possible for any given human to learn any given language. This question was originally a source of inspiration and perhaps one of the primary motivations for UG, but it is this question that actually limits UG to a great extent: if UG is indeed universal, then it must not be too specific. Variability in existing languages worldwide limits UG's potential makeup, while UG in turn theoretically limits the variability of possible languages. To take an extreme example, if UG was just a collection of rules that generated English sentences (like, for example, the Toy Grammar in Table 1), French babies would have a difficult time learning French. Their brains would be hard wired, as it were, to learn English. Indeed, the French language would not exist, because it is the children in a linguistic community who carry the language to the next generation. [Deacon 97]; [Pinker 95]

Deacon makes an interesting point with respect to language processing that is particularly applicable here. Some words, Deacon claims ([Deacon 97], p. 293), serve as syntactic markers in the linear stream of language. These markers help indicate to a listener how to break up a sentence so that the essentially serial data stream can be passed to several small parallel processing units for more efficient comprehension. This process in reverse might be a method for sentence generation, with several relatively independent brain areas contributing, for example, noun phrases or verb phrases, and the junctions between the units providing helpful connective words that have little meaning on their own, like the English word of. [Pinker 95]

Such a serial-to-parallel processing method has interesting implications for neural network modeling that cannot be explored in this paper, but it also has practical applications for students of a foreign language. For example, individuals who learn a foreign language fluently after their critical language acquisition window has closed are generally conscious of their own brain state when speaking the foreign language—even if that consciousness is only a thought about how difficult speaking the foreign language is. Such a conscious awareness of language development skills is interesting from a learning perspective. For example, conscious students of a foreign language often experience a crucial moment when they suddenly feel as though they understand the language on a new level. This might be due to some sort of brain reorganization that facilitates the location of marker words, thus allowing the brain to break up sentences for more efficient understanding.

Developing language

One of the largest problems with Universal Grammar comes when one remembers that humans (and all other organisms) are not built at random; instead, they carry with them a set of information that is necessary for constructing the proteins that either make up or build the rest of the body. This information is stored in chemical form as DNA, which is somewhat segmented into information-carrying packets (genes). Although much work has been done with genetics in recent years, researchers have yet to locate a gene that is responsible for language development.

Indeed, there are several reasons to suspect that no such gene exists. For example, if such a gene did exist, we would expect to find some small proportion of unfortunate individuals who either lack a copy of the gene or have some sort of defunct mutation of it. As far as researchers know, every human with a normal brain has an equal chance of learning to use any human language. Indeed, such language learning almost always takes place at a surprisingly young age [Pinker 95], so a language gene would need to express itself quite early in life. Another reason to suspect that a language gene does not exist is evolutionary: it would be rare to find a process of natural selection that would use up valuable genetic material with an entire gene for language, which is an ability that does not grant immediate success to an individual (like having legs or arms grants success for humans). Additionally, if there were a language gene, one might expect the rest of the human brain to have genes determining its structure. But this is ridiculous: as previously mentioned, the number of neurons and inter-neuronal connections in an adult human's brain is far too great to possibly be genetically determined. Finally, if there were a gene for language ability, one might imagine that it could be implanted in a chimpanzee fetus to get a talking ape, or that it might develop spontaneously in one of today's primates or even mammals. Such occurrences are unknown.

Genes and brain development

Deacon presents several arguments on this subject from developmental psychology and genetics perspectives. For example, brains as a whole are indeed controlled genetically, but only on a very large scale: that of the entire organism. Special genes referred to as homeotic genes express themselves at specific stages during any given organism's early development and govern the gross outline of the organism. The most common of these genes are called homeobox genes. For mammals, homeobox gene expression occurs while the organism is still a fetus; these genes express themselves in certain ways down the length of the body tube and determine what sort of thing the local segment of the body tube will become. For example, the reason that so many vertebrates share the same general body plan (head, torso, appendages) is that most vertebrates share many of the homeobox genes. Thus, early in development, organisms have already determined which parts of the body will be the forebrain, midbrain, and head; which part will be the brain stem and neck; which part will be the spinal cord and torso; and so on. Genetic involvement in the structure and layout of the brain practically stops at this level: there are no genes that govern specific neuron migrations in the cortex, nor are there genes that govern specific wiring configurations. With a system that is composed of billions of independent elements, coding for their configuration genetically would take enormous amounts of DNA and hence be subject to catastrophic failures due to misinterpretation of the DNA or to mutations. ([Deacon 97], pp. 174–181)

Broca's and Wernicke's areas

In addition to lacking a clear genetic basis, it turns out that Universal Grammar is difficult to locate in a particular structure of the brain. Many psychology students learn that there are two major brain areas that are responsible for language processing: Broca's area controls sentence generation, while Wernicke's area is mostly responsible for sentence comprehension and word retrieval. [Kalat 98];[Matlin 98] While this is, for the most part, a true statement, it is a sweeping one, and there are some finer points that should be presented. It is important to keep in mind that the identification of these brain areas as language centers comes mostly from stroke patients, or more recently from PET scans and MRI's. These data, then, are on an enormous scale compared with the scale of an individual neuron or even a small network of such neurons. The picture of brain operation that researchers have is not in focus yet.

Deacon proposes that these two brain areas are indeed located centrally in language processing, but with two conditions. First, being located centrally does not imply sole responsibility. For example, Deacon points out that when stroke patients suffer damage to one of these brain areas, some tissue from surrounding brain areas is damaged as well. It is possible that the brain matter surrounding Broca's and Wernicke's areas has just as much language processing responsibility as the areas themselves. Second, language processing is not located in Broca's and Wernicke's areas in every individual. A larger portion of left-handers have bilateral language control than right-handers, who have predominant language control located in the left hemisphere. This makes it difficult for some left-handers to manipulate objects with either hand while processing language, while right-handers generally have no problem manipulating their left hands while processing language. [Pinker 95] Deacon also presents some evidence that the brain stores and processes a second language in a different location (sometimes even a different hemisphere). This would greatly limit the possible brain locations for UG, if there is a location at all. [Deacon 97]

Additionally, new studies involving more precise brain measurement techniques are slowly revealing more exact language processing locations in the brain. Deacon ([Deacon 97], pp. 288–300) presents the results of several studies using PET scans, rCBF tests, open-brain stimulation experiments, and fMRI's that all indicate just how dispersed language processing in the brain really is. Certainly, Broca's and Wernicke's areas are involved in generating and understanding language, respectively, but each area seems to excite the other, and most language tasks recruit help from other (neighboring and even occasionally distant) areas of the brain. It seems important to keep in mind, in particular, the role of the facial neurons (which sense the position of the tongue and mouth) and the motor neurons that are responsible for speaking.

A Language window

Another argument constraining the potential nature of Universal Grammar comes from interesting studies of children who were deprived of language for whatever reason during the early period of their lives (e.g., [Pinker 95], pp. 291–293). Deaf children seemed particularly vulnerable to this a few decades ago due to medical ignorance of diagnosing deafness at an early age. Children who do not learn any language during the first six or so years of life have considerable difficulties learning any language at all later on. [Pinker 95] If Universal Grammar is present in all individuals, then, it must only be present for a certain critical time period early in life and then get discarded if no language is learned. UG is thus limited in its life span in a certain sense.

Evolving language

Deacon's research actually concentrates on the evolution and use of language in humans, and he uses modern techniques like PET scans to obtain images of brain use while participants are actually performing language tasks. Some of his evolutionary theories have a solid empirical base and present good questions that UG finds it difficult to answer.

Deacon ([Deacon 97], pp. 71–86) begins his arguments with his definitions of a symbolic system; his ideas are clearly influenced by Saussure's work with symbols and linguistic meaning, and also by the work of American philosopher Charles Sanders Pierce, but Deacon obfuscates some of the more standard philosophical terms with definitions of his own. [Hurford 98] For Deacon, these definitions have a further ramification with respect to brain use and learning. Specifically, Deacon proposes that an organism's brain must have a large number of neurons not committed to basic sensorimotor tasks in order to be capable of learning what he calls a symbolic system; that is, an organism must have a large brain to be able to use symbols.

Having a large brain is not enough, though: elephants and whales have enormous brains but are not generally capable of symbolic communication using language. The other essential part of Deacon's proposed puzzle is that an organism's brain must be large and available for symbolic processing. Humans fit the bill because a human's body is large, and hence its brain is large, but a human's brain is absurdly large for its body size compared to other primates. This means humans' brains are only partially devoted to body control (as opposed to a rat's or a monkey's brain, for example), so we have the free neurons necessary to grasp symbolic reference. ([Deacon 97], pp. 168–172)

Conclusion

Given all these data—the vast majority of which were collected after the 1960's—there is one remaining question: how can Universal Grammar still exist in humans? UG is constrained by several factors, so can it still exist? If it can, UG must be willing to admit defeat in some areas.

• Universal Grammar must not be too precise about specific grammars for language learning, or some languages would be unlearnable. Admittedly, this is not a very harsh condition: the variety of human languages is, on one hand, stunning and displays incredible variability, but, on the other hand, there are still many universals when speaking about natural languages. [Pinker 95]

• There is not a gene for Universal Grammar, nor are there genes for specific wiring patterns in developing human brains (or any other brains, for that matter). UG must be a nongenetic phenomenon, so it must come from experience, as Chomsky proposed.

• There is not a specific brain region that is completely devoted to language, or to Universal Grammar. Broca's and Wernicke's areas certainly are involved in most language processing tasks in most people, but there are individuals who process language with both hemispheres, with just the right hemisphere, or with just the left hemisphere. In addition, learning a foreign language seems to recruit different brain areas, not just the ones already in use for the primary language. UG must not be located in one particular brain area, because not every individual processes language with the same brain areas.

• Universal Grammar must not be constant in all individuals across time, since there is some sort of critical language acquisition window.

These are compelling restrictions for Universal Grammar. One might ask if anything at all is left of UG, but there are some characteristics that remain. In particular, one concept that seems to be universal across researchers so far is that all humans have the ability to learn to manipulate a symbolic system in certain ways. [Hurford 98]; [Deacon 97] These certain ways might be what supporters of UG refer to as grammar, or they might be what neurologists and even computer scientists see as the inherent nature of neural network processing methodologies.

In any case, Universal Grammar remains a valid concept today, in some sense. Pursuing the concepts that remain uncontradicted might lead linguistics in new directions.

Footnotes

... language.^1

Saussure claims this is because the linguistic sign is arbitrary, so a linguistic community assigns its own values to signs rather than accepting some preassigned meaning for words. (p. 115)