The Faculty of Language: What Is It, Who Has It, and How Did It Evolve?
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
SCIENCE’S COMPASS ● REVIEW
REVIEW: NEUROSCIENCE
The Faculty of Language: What Is It, Who Has
It, and How Did It Evolve?
Marc D. Hauser,1* Noam Chomsky,2 W. Tecumseh Fitch1
question of language evolution, and of how
We argue that an understanding of the faculty of language requires substantial humans acquired the faculty of language.
interdisciplinary cooperation. We suggest how current developments in linguistics can In exploring the problem of language evo-
be profitably wedded to work in evolutionary biology, anthropology, psychology, and lution, it is important to distinguish between
neuroscience. We submit that a distinction should be made between the faculty of questions concerning language as a commu-
language in the broad sense (FLB) and in the narrow sense (FLN). FLB includes a nicative system and questions concerning the
sensory-motor system, a conceptual-intentional system, and the computational computations underlying this system, such as
mechanisms for recursion, providing the capacity to generate an infinite range of those underlying recursion. As we argue be-
expressions from a finite set of elements. We hypothesize that FLN only includes low, many acrimonious debates in this field
recursion and is the only uniquely human component of the faculty of language. We have been launched by a failure to distinguish
further argue that FLN may have evolved for reasons other than language, hence between these problems. According to one
comparative studies might look for evidence of such computations outside of the view (1), questions concerning abstract com-
domain of communication (for example, number, navigation, and social relations). putational mechanisms are distinct from
those concerning communication, the latter
targeted at problems at the interface between
I
f a martian graced our planet, it would be tures; it might further note that the human abstract computation and both sensory-motor
struck by one remarkable similarity among faculty of language appears to be organized and conceptual-intentional interfaces. This
Earth’s living creatures and a key difference. like the genetic code— hierarchical, genera- view should not, of course, be taken as a
Concerning similarity, it would note that all tive, recursive, and virtually limitless with claim against a relationship between compu-
living things are de-
signed on the basis of
highly conserved de-
velopmental systems
that read an (almost)
universal language en-
coded in DNA base
pairs. As such, life is
arranged hierarchical-
ly with a foundation
of discrete, unblend-
able units (codons, and,
for the most part,
genes) capable of com-
bining to create increas-
ingly complex and vir-
tually limitless varieties
of both species and in-
dividual organisms. In
contrast, it would notice
the absence of a univer-
sal code of communi-
cation (Fig. 1).
If our martian nat- Fig. 1. The animal kingdom has been designed on the basis of highly conserved developmental systems that read an almost
universal language coded in DNA base pairs. This system is shown on the left in terms of a phylogenetic tree. In contrast, animals
uralist were meticu- lack a common universal code of communication, indicated on the right by unconnected animal groups. [Illustration: John Yanson]
lous, it might note
that the faculty medi-
ating human communication appears remark- respect to its scope of expression. With these tation and communication. It is possible, as
ably different from that of other living crea- pieces in hand, this martian might begin to we discuss below, that key computational
wonder how the genetic code changed in such capacities evolved for reasons other than
a way as to generate a vast number of mutu- communication but, after they proved to have
1
Department of Psychology, Harvard University, ally incomprehensible communication sys- utility in communication, were altered be-
Cambridge, MA 02138, USA. Department of Linguis-
2
tics and Philosophy, Massachusetts Institute of Tech-
tems across species while maintaining clarity cause of constraints imposed at both the pe-
nology, Cambridge, MA 02138, USA. of comprehension within a given species. The riphery (e.g., what we can hear and say or see
*To whom correspondence should be addressed. E- martian would have stumbled onto some of and sign, the rapidity with which the auditory
mail: mdhauser@wjh.harvard.edu the essential problems surrounding the cortex can process rapid temporal and spec-
www.sciencemag.org SCIENCE VOL 298 22 NOVEMBER 2002 1569
SCIENCE’S COMPASS
faculty, with a special focus on
comparative work with non-
human animals, and conclude
with a discussion of how in-
quiry might profitably advance,
highlighting some outstanding
problems.
We make no attempt to be
comprehensive in our coverage of
relevant or interesting topics and
problems. Nor is it our goal to
review the history of the field.
Rather, we focus on topics that
make important contact between
empirical data and theoretical po-
sitions about the nature of the lan-
guage faculty. We believe that if
explorations into the problem of
language evolution are to progress,
we need a clear explication of the
computational requirements for
language, the role of evolutionary
theory in testing hypotheses of
character evolution, and a research
program that will enable a produc-
tive interchange between linguists
Fig. 2. A schematic representation of organism-external and -internal factors related to the faculty of language.
and biologists.
FLB includes sensory-motor, conceptual-intentional, and other possible systems (which we leave open); FLN
includes the core grammatical computations that we suggest are limited to recursion. See text for more
complete discussion.
Defining the Target: Two
Senses of the Faculty of
Language
tral changes) and more central levels (e.g., independent questions, leading to a wide The word “language” has highly divergent
conceptual and cognitive structures, pragmat- variety of divergent viewpoints on the evo- meanings in different contexts and disci-
ics, memory limitations). lution of language in the current literature. plines. In informal usage, a language is un-
At least three theoretical issues cross-cut There is, however, an emerging consensus derstood as a culturally specific communica-
the debate on language evolution. One of the that, although humans and animals share a tion system (English, Navajo, etc.). In the
oldest problems among theorists is the diversity of important computational and varieties of modern linguistics that concern
“shared versus unique” distinction. Most cur- perceptual resources, there has been sub- us here, the term “language” is used quite
rent commentators agree that, although bees stantial evolutionary remodeling since we differently to refer to an internal component
dance, birds sing, and chimpanzees grunt, diverged from a common ancestor some 6 of the mind/brain (sometimes called “internal
these systems of communication differ qual- million years ago. The empirical challenge language” or “I-language”). We assume that
itatively from human language. In particular, is to determine what was inherited un- this is the primary object of interest for the
animal communication systems lack the rich changed from this common ancestor, what study of the evolution and function of the
expressive and open-ended power of human has been subjected to minor modifications, language faculty. However, this biologically
language (based on humans’ capacity for re- and what (if anything) is qualitatively new. and individually grounded usage still leaves
cursion). The evolutionary puzzle, therefore, The additional evolutionary challenge is to much open to interpretation (and misunder-
lies in working out how we got from there to determine what selectional pressures led to standing). For example, a neuroscientist
here, given this apparent discontinuity. A sec- adaptive changes over time and to under- might ask: What components of the human
ond issue revolves around whether the evo- stand the various constraints that channeled nervous system are recruited in the use of
lution of language was gradual versus salta- this evolutionary process. Answering these language in its broadest sense? Because any
tional; this differs from the first issue because questions requires a collaborative effort aspect of cognition appears to be, at least in
a qualitative discontinuity between extant among linguists, biologists, psychologists, principle, accessible to language, the broadest
species could have evolved gradually, involv- and anthropologists. answer to this question is, probably, “most of
ing no discontinuities during human evolu- One aim of this essay is to promote a it.” Even aspects of emotion or cognition not
tion. Finally, the “continuity versus exapta- stronger connection between biology and readily verbalized may be influenced by lin-
tion” issue revolves around the problem of linguistics by identifying points of contact guistically based thought processes. Thus,
whether human language evolved by gradual and agreement between the fields. Al- this conception is too broad to be of much
extension of preexisting communication sys- though this interdisciplinary marriage was use. We therefore delineate two more restrict-
tems, or whether important aspects of lan- inaugurated more than 50 years ago, it has ed conceptions of the faculty of language, one
guage have been exapted away from their not yet been fully consummated. We hope broader and more inclusive, the other more
previous adaptive function (e.g., spatial or to further this goal by, first, helping to restricted and narrow (Fig. 2).
numerical reasoning, Machiavellian social clarify the biolinguistic perspective on lan- Faculty of language— broad sense
scheming, tool-making). guage and its evolution (2–7). We then (FLB). FLB includes an internal computa-
Researchers have adopted extreme or in- review some promising empirical ap- tional system (FLN, below) combined with
termediate positions regarding these basically proaches to the evolution of the language at least two other organism-internal sys-
1570 22 NOVEMBER 2002 VOL 298 SCIENCE www.sciencemag.orgSCIENCE’S COMPASS
tems, which we call “sensory-motor” and ly, a system of sound-meaning connections; arbitrary upper bound to sentence length. In
“conceptual-intentional.” Despite debate on the potential infiniteness of this system has these respects, language is directly analogous
the precise nature of these systems, and been explicitly recognized by Galileo, Des- to the natural numbers (see below).
about whether they are substantially shared cartes, and the 17th-century “philosophical At a minimum, then, FLN includes the ca-
with other vertebrates or uniquely adapted grammarians” and their successors, notably pacity of recursion. There are many organism-
to the exigencies of language, we take as von Humboldt. One goal of the study of FLN internal factors, outside FLN or FLB, that im-
uncontroversial the existence of some bio- and, more broadly, FLB is to discover just pose practical limits on the usage of the system.
logical capacity of humans that allows us how the faculty of language satisfies these For example, lung capacity imposes limits on
(and not, for example, chimpanzees) to basic and essential conditions. the length of actual spoken sentences, whereas
readily master any human language without The core property of discrete infinity is working memory imposes limits on the com-
explicit instruction. FLB includes this ca- intuitively familiar to every language user. plexity of sentences if they are to be under-
pacity, but excludes other organism- Sentences are built up of discrete units: There standable. Other limitations—for example, on
internal systems that are necessary but not are 6-word sentences and 7-word sentences, concept formation or motor output speed—
sufficient for language (e.g., memory, res- but no 6.5-word sentences. There is no long- represent aspects of FLB, which have their own
piration, digestion, circulation, etc.). est sentence (any candidate sentence can be evolutionary histories and may have played a
Faculty of language—narrow sense trumped by, for example, embedding it in role in the evolution of the capacities of FLN.
(FLN). FLN is the abstract linguistic compu- “Mary thinks that . . .”), and there is no non- Nonetheless, one can profitably inquire into the
tational system alone, independent of the oth- evolution of FLN without an
er systems with which it interacts and inter- immediate concern for these
faces. FLN is a component of FLB, and the limiting aspects of FLB. This
mechanisms underlying it are some subset of is made clear by the observa-
those underlying FLB. tion that, although many
Others have agreed on the need for a aspects of FLB are shared
restricted sense of “language” but have sug- with other vertebrates, the
gested different delineations. For example, core recursive aspect of FLN
Liberman and his associates (8) have argued currently appears to lack any
that the sensory-motor systems were specifi- analog in animal communi-
cally adapted for language, and hence should cation and possibly other do-
be considered part of FLN. There is also a mains as well. This point,
long tradition holding that the conceptual- therefore, represents the
intentional systems are an intrinsic part of deepest challenge for a com-
language in a narrow sense. In this article, we parative evolutionary ap-
leave these questions open, restricting atten- proach to language. We be-
tion to FLN as just defined but leaving the lieve that investigations of
possibility of a more inclusive definition this capacity should include
open to further empirical research. domains other than commu-
The internal architecture of FLN, so con- nication (e.g., number, social
ceived, is a topic of much current research relationships, navigation).
and debate (4). Without prejudging the is- Given the distinctions
sues, we will, for concreteness, adopt a par- between FLB and FLN and
ticular conception of this architecture. We the theoretical distinctions
assume, putting aside the precise mecha- raised above, we can define
nisms, that a key component of FLN is a a research space as sketched
computational system (narrow syntax) that in Fig. 3. This research
generates internal representations and maps space identifies, as viable,
them into the sensory-motor interface by the problems concerning the
phonological system, and into the conceptu- evolution of sensory-motor
al-intentional interface by the (formal) se- systems, of conceptual-in-
mantic system; adopting alternatives that tentional systems, and of
have been proposed would not materially FLN. The comparative ap-
modify the ensuing discussion. All approach- proach, to which we turn
es agree that a core property of FLN is recur- next, provides a framework
sion, attributed to narrow syntax in the con- for addressing questions
ception just outlined. FLN takes a finite set of about each of these com-
elements and yields a potentially infinite ar- ponents of the faculty of
ray of discrete expressions. This capacity of language.
FLN yields discrete infinity (a property that
also characterizes the natural numbers). Each The Comparative
of these discrete expressions is then passed to Approach to Language
the sensory-motor and conceptual-intentional Fig. 3. Investigations into the evolution of the faculty of language Evolution
systems, which process and elaborate this are confronted with a three-dimensional research space that The empirical study of the
includes three comparative-evolutionary problems cross-cut by
information in the use of language. Each the core components of the faculty of language. Thus, for each
evolution of language is be-
expression is, in this sense, a pairing of sound problem, researchers can investigate details of the sensory-motor set with difficulties. Lin-
and meaning. It has been recognized for thou- system, the conceptual-intentional system, FLN, and the interfac- guistic behavior does not
sands of years that language is, fundamental- es among these systems. fossilize, and a long tradi-
www.sciencemag.org SCIENCE VOL 298 22 NOVEMBER 2002 1571SCIENCE’S COMPASS
tion of analysis of fossil skull shape and course, that a trait evolved in nonhuman an- ponents share a deeply conserved neural and
cranial endocasts has led to little consensus imals and humans independently, as analogs developmental foundation: Most aspects of
about the evolution of language (7, 9). A rather than homologs. This would preserve neurophysiology and development—including
more tractable and, we think, powerful ap- the possibility that the trait evolved for lan- regulatory and structural genes, as well as neu-
proach to problems of language evolution is guage in humans but evolved for some other ron types and neurotransmitters—are shared
provided by the comparative method, which reason in the comparative animal group. In among vertebrates. That such close parallels
uses empirical data from living species to cases where the comparative group is a non- have evolved suggests the existence of impor-
draw detailed inferences about extinct ances- human primate, and perhaps especially chim- tant constraints on how vertebrate brains can
tors (3, 10 –12). The comparative method was panzees, the plausibility of this evolutionary acquire large vocabularies of complex, learned
the primary tool used by Darwin (13, 14) to scenario is weaker. In any case, comparative sounds. Such constraints may essentially force
analyze evolutionary phenomena and contin- data are critical to this judgment. natural selection to come up with the same
ues to play a central role throughout modern Despite the crucial role of homology in solution repeatedly when confronted with sim-
evolutionary biology. Although scholars in- comparative biology, homologous traits are not ilar problems.
terested in language evolution have often ig- the only relevant source of evolutionary data.
nored comparative data altogether or focused The convergent evolution of similar characters Testing Hypotheses About the
narrowly on data from nonhuman primates, in two independent clades, termed “analogies” Evolution of the Faculty of Language
current thinking in neuroscience, molecular or “homoplasies,” can be equally revealing Given the definitions of the faculty of lan-
biology, and developmental biology indicates (20). The remarkably similar (but nonhomolo- guage, together with the comparative frame-
that many aspects of neural and developmen- gous) structures of human and octopus eyes work, we can distinguish several plausible
tal function are highly conserved, encourag- reveal the stringent constraints placed by the hypotheses about the evolution of its various
ing the extension of the comparative method laws of optics and the contingencies of devel- components. Here, we suggest two hypothe-
to all vertebrates (and perhaps beyond). For opment on an organ capable of focusing a sharp ses that span the diversity of opinion among
several reasons, detailed below, we believe image onto a sheet of receptors. Detailed anal- current scholars, plus a third of our own.
that the comparative method should play a ogies between the parts of the vertebrate and Hypothesis 1: FLB is strictly homologous
more central role in future discussions of cephalopod eye also provide independent evi- to animal communication. This hypothesis
language evolution. dence that each component is an adaptation for holds that homologs of FLB, including FLN,
An overarching concern in studies of lan- image formation, shaped by natural selection. exist ( perhaps in less developed or otherwise
guage evolution is with whether particular Furthermore, the discovery that remarkably modified form) in nonhuman animals (3, 10,
components of the faculty of language conservative genetic cascades underlie the de- 26). This has historically been a popular hy-
evolved specifically for human language and, velopment of such analogous structures pro- pothesis outside of linguistics and closely
therefore (by extension), are unique to hu- vides important insights into the ways in allied fields, and has been defended by some
mans. Logically, the human uniqueness claim which developmental mechanisms can in the speech sciences. According to this
must be based on data indicating an absence channel evolution (21). Thus, although po- hypothesis, human FLB is composed of the
of the trait in nonhuman animals and, to be tentially misleading for taxonomists, anal- same functional components that underlie
taken seriously, requires a substantial body of ogies provide critical data about adaptation communication in other species.
relevant comparative data. More concretely, under physical and developmental con- Hypothesis 2: FLB is a derived, uniquely
if the language evolution researcher wishes to straints. Casting the comparative net more human adaptation for language. According
make the claim that a trait evolved uniquely broadly, therefore, will most likely reveal to this hypothesis, FLB is a highly complex
in humans for the function of language pro- larger regularities in evolution, helping to adaptation for language, on a par with the
cessing, data indicating that no other animal address the role of such constraints in the vertebrate eye, and many of its core compo-
has this particular trait are required. evolution of language. nents can be viewed as individual traits that
Although this line of reasoning may ap- An analogy recognized as particularly rele- have been subjected to selection and perfect-
pear obvious, it is surprisingly common for a vant to language is the acquisition of song by ed in recent human evolutionary history. This
trait to be held up as uniquely human before birds (12). In contrast to nonhuman primates, appears to represent the null hypothesis for
any appropriate comparative data are avail- where the production of species-typical vocal- many scholars who take the complexity of
able. A famous example is categorical per- izations is largely innate (22), most songbirds language seriously (27, 28). The argument
ception, which when discovered seemed so learn their species-specific song by listening to starts with the assumption that FLB, as a
finely tuned to the details of human speech as conspecifics, and they develop highly aberrant whole, is highly complex, serves the function
to constitute a unique human adaptation (15, song if deprived of such experience. Current of communication with admirable effective-
16). It was some time before the same under- investigation of birdsong reveals detailed and ness, and has an ineliminable genetic compo-
lying perceptual discontinuities were discov- intriguing parallels with speech (11, 23, 24). nent. Because natural selection is the only
ered in chinchillas and macaques (17, 18), For instance, many songbirds pass through a known biological mechanism capable of gen-
and even birds (19), leading to the opposite critical period in development beyond which erating such functional complexes [the argu-
conclusion that the perceptual basis for cate- they produce defective songs that no amount of ment from design (29)], proponents of this
gorical perception is a primitive vertebrate acoustic input can remedy, reminiscent of the view conclude that natural selection has
characteristic that evolved for general audito- difficulty adult humans have in fully mastering played a powerful role in shaping many as-
ry processing, as opposed to specific speech new languages. Further, and in parallel with the pects of FLB, including FLN, and, further,
processing. Thus, a basic and logically in- babbling phase of vocalizing or signing human that many of these are without parallel in
eliminable role for comparative research on infants (25), young birds pass through a phase nonhuman animals. Although homologous
language evolution is this simple and essen- of song development in which they spontane- mechanisms may exist in other animals, the
tially negative one: A trait present in nonhu- ously produce amorphous versions of adult human versions have been modified by nat-
man animals did not evolve specifically for song, termed “subsong” or “babbling.” Al- ural selection to the extent that they can be
human language, although it may be part of though the mechanisms underlying the acquisi- reasonably seen as constituting novel traits,
the language faculty and play an intimate role tion of birdsong and human language are clear- perhaps exapted from other contexts [e.g.,
in language processing. It is possible, of ly analogs and not homologs, their core com- social intelligence, tool-making (7, 30 –32)].
1572 22 NOVEMBER 2002 VOL 298 SCIENCE www.sciencemag.orgSCIENCE’S COMPASS
Hypothesis 3: Only FLN is uniquely human. they appear in narrow syntax and the mappings all peripheral components of FLB are shared
On the basis of data reviewed below, we hy- to the interfaces. If FLN is indeed this restrict- with other animals, in more or less the same
pothesize that most, if not all, of FLB is based ed, this hypothesis has the interesting effect of form as they exist in humans, with differences
on mechanisms shared with nonhuman animals nullifying the argument from design, and thus of quantity rather than kind (9, 34). What is
(as held by hypothesis 1). In contrast, we sug- rendering the status of FLN as an adaptation unique to our species is quite specific to FLN,
gest that FLN—the computational mechanism open to question. Proponents of the idea that and includes its internal operations as well as its
of recursion—is recently evolved and unique to FLN is an adaptation would thus need to supply interface with the other organism-internal sys-
our species (33, 34). According to this hypoth- additional data or arguments to support this tems of FLB.
esis, much of the complexity manifested in viewpoint. Each of these hypotheses is plausible to
language derives from complexity in the pe- The available comparative data on animal some degree. Ultimately, they can be distin-
ripheral components of FLB, especially those communication systems suggest that the faculty guished only by empirical data, much of which
underlying the sensory-motor (speech or sign) of language as a whole relies on some uniquely is currently unavailable. Before reviewing some
and conceptual-intentional interfaces, com- human capacities that have evolved recently in of the relevant data, we briefly consider some
bined with sociocultural and communicative the approximately 6 million years since our key distinctions between them. From a compar-
contingencies. FLB as a whole thus has an divergence from a chimpanzee-like common ative evolutionary viewpoint, an important
ancient evolutionary history, long predating the ancestor (35). Hypothesis 3, in its strongest question is whether linguistic precursors were
emergence of language, and a comparative form, suggests that only FLN falls into this involved in communication or in something
analysis is necessary to understand this com- category (34). By this hypothesis, FLB contains else. Proponents of both hypotheses 1 and 2
plex system. By contrast, according to recent a wide variety of cognitive and perceptual posit a direct correspondence, by descent with
linguistic theory, the computations underlying mechanisms shared with other species, but only modification, between some trait involved in
FLN may be quite limited. In fact, we propose those mechanisms underlying FLN—particu- FLB in humans and a similar trait in another
in this hypothesis that FLN comprises only the larly its capacity for discrete infinity—are species; these hypotheses differ in whether
core computational mechanisms of recursion as uniquely human. This hypothesis suggests that the precursors functioned in communication.
Table 1. A sampler of empirical approaches to understanding the evolution of the faculty of language, including both broad (FLB) and narrow (FLN)
components.
Empirical problem Examples References
FLB—sensory-motor system
Vocal imitation and invention Tutoring studies of songbirds, analyses of vocal dialects in whales, spontaneous imitation (11, 12, 24, 65)
of artificially created sounds in dolphins
Neurophysiology of Studies assessing whether mirror neurons, which provide a core substrate for the (67, 68, 71)
action-perception systems action-perception system, may subserve gestural and ( possibly) vocal imitation
Discriminating the sound patterns Operant conditioning studies of the prototype magnet effect in macaques and starlings (52, 120)
of language
Constraints imposed by vocal tract Studies of vocal tract length and formant dispersion in birds and primates (54–61)
anatomy
Biomechanics of sound production Studies of primate vocal production, including the role of mandibular oscillations (121, 122)
Modalities of language production Cross-modal perception and sign language in humans versus unimodal communication in (3, 25, 123)
and perception animals
FLB—conceptual-intentional system
Theory of mind, attribution of Studies of the seeing/knowing distinction in chimpanzees (84, 86–89)
mental states
Capacity to acquire nonlinguistic Studies of rhesus monkeys and the object/kind concept (10, 76, 77, 124)
conceptual representations
Referential vocal signals Studies of primate vocalizations used to designate predators, food, and social (3, 78, 90, 91, 93,
relationships 94, 97)
Imitation as a rational, intentional Comparative studies of chimpanzees and human infants suggesting that only the latter (125–127)
system read intentionality into action, and thus extract unobserved rational intent
Voluntary control over signal Comparative studies that explore the relationship between signal production and the (3, 10, 92, 128)
production as evidence of composition of a social audience
intentional communication
FLN—recursion
Spontaneous and training methods Studies of serial order learning and finite-state grammars in tamarins and macaques (114, 116, 117,
designed to uncover constraints 129)
on rule learning
Sign or artificial language in Studies exploring symbol sequencing and open-ended combinatorial manipulation (130, 131)
trained apes and dolphins
Models of the faculty of language Game theory models of language acquisition, reference, and universal grammar (72–74)
that attempt to uncover the
necessary and sufficient
mechanisms
Experiments with animals that Operant conditioning studies to determine whether nonhuman primates can represent (102–106, 132)
explore the nature and content number, including properties such as ordinality and cardinality, using such
of number representation representations in conjunction with mathematical operands (e.g., add, divide)
Shared mechanisms across Evolution of musical processing and structure, including analyses of brain function and (133–135)
different cognitive domains comparative studies of music perception
www.sciencemag.org SCIENCE VOL 298 22 NOVEMBER 2002 1573SCIENCE’S COMPASS
Although many aspects of FLB very likely leaves high in the air. Some aspects of such human mechanisms special to speech, despite a
arose in this manner, the important issue for trees are clearly adaptations channeled by persistent tendency to assume uniqueness even
these hypotheses is whether a series of gradual these constraints; others (e.g., the popping of in the absence of relevant animal data.
modifications could lead eventually to the ca- xylem tubes on hot days, the propensity to be On the side of perception, for example,
pacity of language for infinite generativity. De- toppled in hurricanes) are presumably un- many species show an impressive ability to
spite the inarguable existence of a broadly avoidable by-products of such constraints. both discriminate between and generalize
shared base of homologous mechanisms in- Recent work on FLN (4, 41–43) suggests over human speech sounds, using formants as
volved in FLB, minor modifications to this the possibility that at least the narrow-syntactic the critical discriminative cue (17–19, 49 –
foundational system alone seem inadequate to component satisfies conditions of highly effi- 51). These data provide evidence not only of
generate the fundamental difference— discrete cient computation to an extent previously unsus- categorical perception, but also of the ability
infinity— between language and all known pected. Thus, FLN may approximate a kind of to discriminate among prototypical exem-
forms of animal communication. This claim is “optimal solution” to the problem of linking plars of different phonemes (52). Further, in
one of several reasons why we suspect that the sensory-motor and conceptual-intentional the absence of training, nonhuman primates
hypothesis 3 may be a productive way to char- systems. In other words, the generative process- can discriminate sentences from two different
acterize the problem of language evolution. es of the language system may provide a languages on the basis of rhythmic differenc-
A primary issue separating hypotheses 2 near-optimal solution that satisfies the interface es between them (53).
and 3 is whether the uniquely human capac- conditions to FLB. Many of the details of lan- On the side of production, birds and non-
ities of FLN constitute an adaptation. The guage that are the traditional focus of linguistic human primates naturally produce and per-
viewpoint stated in hypothesis 2, especially study [e.g., subjacency, Wh- movement, the ceive formants in their own species-typical
the notion that FLN in particular is a highly existence of garden-path sentences (4, 44)] may vocalizations (54 –59). The results also shed
evolved adaptation, has generated much en- represent by-products of this solution, gener- light on discussions of the uniquely human
thusiasm recently [e.g., (36)], especially ated automatically by neural/computational structure of the vocal tract and the unusual
among evolutionary psychologists (37, 38). constraints and the structure of FLB — descended larynx of our species (7, 48, 60),
At present, however, we see little reason to components that lie outside of FLN. Even because new evidence shows that several oth-
believe either that FLN can be anatomized novel capacities such as recursion are imple- er mammalian species also have a descended
into many independent but interacting traits, mented in the same type of neural tissue as the larynx (61). Because these nonhuman species
each with its own independent evolutionary rest of the brain and are thus constrained by lack speech, a descended larynx clearly has
history, or that each of these traits could have biophysical, developmental, and computation- nonphonetic functions; one possibility is ex-
been strongly shaped by natural selection, al factors shared with other vertebrates. Hy- aggerating apparent size. Although this par-
given their tenuous connection to communi- pothesis 3 raises the possibility that structural ticular anatomical modification undoubtedly
cative efficacy (the surface or phenotypic details of FLN may result from such preexisting plays an important role in speech production
function upon which selection presumably constraints, rather than from direct shaping by in modern humans, it need not have first
acted). natural selection targeted specifically at com- evolved for this function. The descended lar-
We consider the possibility that certain spe- munication. Insofar as this proves to be true, ynx may thus be an example of classic Dar-
cific aspects of the faculty of language are such structural details are not, strictly speaking, winian preadaptation.
“spandrels”— by-products of preexisting con- adaptations at all. This hypothesis and the Many phenomena in human speech percep-
straints rather than end products of a history of alternative selectionist account are both viable tion have not yet been investigated in animals
natural selection (39). This possibility, which and can eventually be tested with comparative [e.g., the McGurk effect, an illusion in which
opens the door to other empirical lines of inqui- data. the syllable perceived from a talking head rep-
ry, is perfectly compatible with our firm support resents the interaction between an articulatory
of the adaptationist program. Indeed, it follows Comparative Evidence for the Faculty gesture seen and a different syllable heard; see
directly from the foundational notion that adap- of Language (62)]. However, the available data suggest a
tation is an “onerous concept” to be invoked Study of the evolution of language has accel- much stronger continuity between animals and
only when alternative explanations fail (40). erated in the past decade (45, 46). Here, we humans with respect to speech than previously
The question is not whether FLN in toto is offer a highly selective review of some of believed. We argue that the continuity hypoth-
adaptive. By allowing us to communicate an these studies, emphasizing animal work that esis thus deserves the status of a null hypothe-
endless variety of thoughts, recursion is clearly seems particularly relevant to the hypotheses sis, which must be rejected by comparative
an adaptive computation. The question is advanced above; many omissions were nec- work before any claims of uniqueness can be
whether particular components of the function- essary for reasons of space, and we firmly validated. For now, this null hypothesis of no
ing of FLN are adaptations for language, spe- believe that a broad diversity of methods and truly novel traits in the speech domain appears
cifically acted upon by natural selection—or, perspectives will ultimately provide the rich- to stand.
even more broadly, whether FLN evolved for est answers to the problem of language evo- There is, however, a striking ability tied to
reasons other than communication. lution. For this reason, we present a broader speech that has received insufficient atten-
An analogy may make this distinction sampler of the field’s offerings in Table 1. tion: the human capacity for vocal imitation
clear. The trunk and branches of trees are How “special” is speech? Comparative (63, 64). Imitation is obviously a necessary
near-optimal solutions for providing an indi- study of the sensory-motor system. Starting with component of the human capacity to acquire
vidual tree’s leaves with access to sunlight. early work on speech perception, there has been a shared and arbitrary lexicon, which is itself
For shrubs and small trees, a wide variety of a tradition of considering speech “special,” and central to the language capacity. Thus, the
forms (spreading, spherical, multistalked, thus based on uniquely human mechanisms capacity to imitate was a crucial prerequisite
etc.) provide good solutions to this problem. adapted for speech perception and/or produc- of FLB as a communicative system. Vocal
For a towering rainforest canopy tree, how- tion [e.g., (7, 8, 47, 48)]. This perspective has imitation and learning are not uniquely hu-
ever, most of these forms are rendered im- stimulated a vigorous research program study- man. Rich multimodal imitative capacities
possible by the various constraints imposed ing animal speech perception and, more recent- are seen in other mammals (dolphins) and
by the properties of cellulose and the prob- ly, speech production. Surprisingly, this re- some birds ( parrots), with most songbirds
lems of sucking water and nutrients up to the search has turned up little evidence for uniquely exhibiting a well-developed vocal imitative
1574 22 NOVEMBER 2002 VOL 298 SCIENCE www.sciencemag.orgSCIENCE’S COMPASS
capacity (65). What is surprising is that mon- particular way and when the same individual have rich conceptual representations (76, 77).
keys show almost no evidence of visually sees someone else act in this same way (67, Surprisingly, however, there is a mismatch be-
mediated imitation, with chimpanzees show- 68), these neurons are not sufficient for imi- tween the conceptual capacities of animals and
ing only slightly better capacities (66). Even tation in macaques, as many have presumed: the communicative content of their vocal and
more striking is the virtual absence of evi- As mentioned, there is no convincing evi- visual signals (78, 79). For example, although a
dence for vocal imitation in either monkeys dence of vocal or visual imitation in mon- wide variety of nonhuman primates have access
or apes (3). For example, intensively trained keys. Consequently, as neuroimaging studies to rich knowledge of who is related to whom, as
chimpanzees are incapable of acquiring any- continue to explore the neural basis of imita- well as who is dominant and who is subordi-
thing but a few poorly articulated spoken tion in humans (69 –71), it will be important nate, their vocalizations only coarsely express
words, whereas parrots can readily acquire a to distinguish between the necessary and suf- such complexities.
large vocal repertoire. With respect to their ficient neural correlates of imitation. This is Studies using classical training approach-
own vocalizations, there are few convincing especially important, given that some recent es as well as methods that tap spontaneous
studies of vocal dialects in primates, thereby attempts to model the evolution of language abilities reveal that animals acquire and use a
suggesting that they lack a vocal imitative begin with a hypothetical organism that is wide range of abstract concepts, including
capacity (3, 65). Evidence for spontaneous equipped with the capacity for imitation and tool, color, geometric relationships, food, and
visuomanual imitation in chimpanzees is not intentionality, as opposed to working out how number (66, 76 – 82). More controversially,
much stronger, although with persistent train- these mechanisms evolved in the first place but of considerable relevance to intentional
ing they can learn several hundred hand [see below; (72–74)]. If a deeper evolution- aspects of language and conditions of felici-
signs. Further, even in cases where
nonhuman animals are capable of im-
itating in one modality (e.g., song
copying in songbirds), only dolphins
and humans appear capable of imita-
tion in multiple modalities. The de-
tachment from modality-specific in-
puts may represent a substantial
change in neural organization, one
that affects not only imitation but
also communication; only humans
can lose one modality (e.g., hear-
ing) and make up for this deficit by
communicating with complete com-
petence in a different modality (i.e.,
signing).
Our discussion of limitations is
not meant to diminish the impressive
achievements of monkeys and apes,
but to highlight how different the
mechanisms underlying the produc-
tion of human and nonhuman primate
gestures, either vocally expressed or
signed, must be. After all, the aver-
age high school graduate knows up to
60,000 words, a vocabulary achieved
with little effort, especially when
contrasted with the herculean efforts
devoted to training animals. In sum,
the impressive ability of any normal
human child for vocal imitation may
represent a novel capacity that
evolved in our recent evolutionary
history, some time after the diver-
gence from our chimpanzee-like an-
cestors. The existence of analogs in Fig. 4. The distribution of imitation in the animal kingdom is patchy. Some animals such as songbirds,
dolphins, and humans have evolved exceptional abilities to imitate; other animals, such as apes and
distantly related species, such as monkeys, either lack such abilities or have them in a relatively impoverished form. [Illustration: John Yanson]
birds and cetaceans, suggests consid-
erable potential for the detailed com-
parative study of vocal imitation. There are, ary exploration is desired, one dating back to tous use, some studies claim that animals
however, potential traps that must be avoid- a chimpanzee-like ancestor, then we need to have a theory of mind (83– 85), including a
ed, especially with respect to explorations of explain how and why such capacities sense of self and the ability to represent the
the neurobiological substrates of imitation. emerged from an ancestral node that lacked beliefs and desires of other group members.
For example, although macaque monkeys such abilities (75) (Fig. 4). On the side of positive support, recent studies
and humans are equipped with so-called The conceptual-intentional systems of non- of chimpanzees suggest that they recognize
“mirror neurons” in the premotor cortex that linguistic animals. A wide variety of studies the perceptual act of seeing as a proxy for the
respond both when an individual acts in a indicate that nonhuman mammals and birds mental state of knowing (84, 86, 87). These
www.sciencemag.org SCIENCE VOL 298 22 NOVEMBER 2002 1575SCIENCE’S COMPASS
studies suggest that at least chimpanzees, but trary in terms of its association with a partic- ular human language as consisting of words and
perhaps no other nonhuman animals, have a ular context, is sufficient to enable listeners to computational procedures (“rules”) for con-
rudimentary theory of mind. On the side of respond appropriately without requiring any structing expressions from them. The computa-
negative support, other studies suggest that other contextual information. Third, the num- tional system has the recursive property briefly
even chimpanzees lack a theory of mind, ber of such signals in the repertoire is small, outlined earlier, which may be a distinct human
failing, for example, to differentiate between restricted to objects and events experienced property. However, key aspects of words may
ignorant and knowledgeable individuals with in the present, with no evidence of creative also be distinctively human. There are, first of
respect to intentional communication (88, production of new sounds for new situations. all, qualitative differences in scale and mode of
89). Because these experiments make use of Fourth, the acoustic morphology of the calls acquisition, which suggest that quite different
different methods and are based on small is fixed, appearing early in development, with mechanisms are involved; as pointed out above,
sample sizes, it is not possible at present to experience only playing a role in refining the there is no evidence for vocal imitation in non-
derive any firm conclusions about the pres- range of objects or events that elicit such human primates, and although human children
ence or absence of mental state attribution in calls. Fifth, there is no evidence that calling is may use domain-general mechanisms to ac-
animals. Independently of how this contro- intentional in the sense of taking into account quire and recall words (98, 99), the rate at
versy is resolved, however, the best evidence what other individuals believe or want. which children build the lexicon is so massively
of referential communication in animals Early interpretations of this work suggest- different from nonhuman primates that one
comes not from chimpanzees but from a va- ed that when animals vocalize, they are func- must entertain the possibility of an indepen-
dently evolved mechanism. Further-
more, unlike the best animal exam-
ples of putatively referential signals,
most of the words of human lan-
guage are not associated with specif-
ic functions (e.g., warning cries, food
announcements) but can be linked to
virtually any concept that humans
can entertain. Such usages are often
highly intricate and detached from
the here and now. Even for the sim-
plest words, there is typically no
straightforward word-thing relation-
ship, if “thing” is to be understood in
mind-independent terms. Without
pursuing the matter here, it appears
that many of the elementary proper-
ties of words—including those that
enter into referentiality— have only
weak analogs or homologs in natural
animal communication systems, with
only slightly better evidence from the
training studies with apes and dol-
phins. Future research must therefore
provide stronger support for the pre-
Fig. 5. Human and nonhuman animals exhibit the capacity to compute numerosities, including small precise cursor position, or it must instead
number quantification and large approximate number estimation. Humans may be unique, however, in the ability
to show open-ended, precise quantificational skills with large numbers, including the integer count list. In parallel
abandon this hypothesis, arguing that
with the faculty of language, our capacity for number relies on a recursive computation. [Illustration: John Yanson] this component of FLB (conceptual-
intentional) is also uniquely human.
Discrete infinity and constraints
riety of monkeys and birds, species for which tionally referring to the objects and events on learning. The data summarized thus far,
there is no convincing evidence for a theory that they have encountered. As such, vervet although far from complete, provide overall
of mind. alarm calls and rhesus monkey food calls, to support for the position of continuity between
The classic studies of vervet monkey take two examples, were interpreted as word- humans and other animals in terms of FLB.
alarm calls (90) have now been joined by like, with callers referring to different kinds However, we have not yet addressed one
several others, each using comparable meth- of predators or different kinds of food. More issue that many regard as lying at the heart of
ods, with extensions to different species (ma- recent discussions have considerably weak- language: its capacity for limitless expressive
caques, Diana monkeys, meerkats, prairie ened this interpretation, suggesting that if the power, captured by the notion of discrete
dogs, chickens) and different communicative signal is referential at all, it is in the mind of infinity. It seems relatively clear, after nearly
contexts (social relationships, food, inter- the listener who can extract information a century of intensive research on animal
group aggression) (91–97). From these stud- about the signaler’s current context from the communication, that no species other than
ies we can derive five key points relevant to acoustic structure of the call alone (78, 95). humans has a comparable capacity to recom-
our analysis of the faculty of language. First, Despite this evidence that animals can extract bine meaningful units into an unlimited vari-
individuals produce acoustically distinctive information from the signal, there are several ety of larger structures, each differing sys-
calls in response to functionally important reasons why additional evidence is required tematically in meaning. However, little
contexts, including the detection of predators before such signals can be considered as pre- progress has been made in identifying the
and the discovery of food. Second, the acous- cursors for, or homologs of, human words. specific capabilities that are lacking in other
tic morphology of the signal, although arbi- Roughly speaking, we can think of a partic- animals.
1576 22 NOVEMBER 2002 VOL 298 SCIENCE www.sciencemag.orgSCIENCE’S COMPASS
The astronomical variety of sentences any around the mean. These results have led to a difference from the training regime of ani-
natural language user can produce and under- the idea that animals, including human in- mals: Children typically first learn an arbi-
stand has an important implication for lan- fants and adults, can represent number ap- trary ordered list of symbols (“1, 2, 3, 4 . . . ”)
guage acquisition, long a core issue in devel- proximately as a magnitude with scalar vari- and later learn the precise meaning of such
opmental psychology. A child is exposed to ability (101, 102). Number discrimination is words; apes and parrots, in contrast, were
only a small proportion of the possible sen- limited in this system by Weber’s law, with taught the meanings one by one without
tences in its language, thus limiting its data- greater discriminability among small num- learning the list. As Carey (103) has argued,
base for constructing a more general version bers than among large numbers (keeping dis- this may represent a fundamental difference
of that language in its own mind/brain. This tances between pairs constant) and between in experience, a hypothesis that could be
point has logical implications for any system numbers that are farther apart (e.g., 7 versus tested by first training animals with an arbi-
that attempts to acquire a natural language on 8 is harder than 7 versus 12). The approxi- trary ordered list.
the basis of limited data. It is immediately mate number sense is accompanied by a sec- A second possible limitation on the class
obvious that given a finite array of data, there ond precise mechanism that is limited to val- of learnable structures concerns the kinds of
are infinitely many theories consistent with it ues less than 4 but accurately distinguishes 1 statistical inferences that animals can com-
but inconsistent with one another. In the from 2, 2 from 3, and 3 from 4; this second pute. Early work in computational linguistics
present case, there are in principle infinitely system appears to be recruited in the context (108 –110) suggested that we can profitably
many target systems ( potential I-languages) of object tracking and is limited by working think about language as a system of rules
consistent with the data of experience, and memory constraints (103). Of direct rele- placed within a hierarchy of increasing com-
unless the search space and acquisition mech- vance to the current discussion, animals can plexity. At the lowest level of the hierarchy
anisms are constrained, selection among be trained to understand the meaning of num- are rule systems that are limited to local
them is impossible. A version of the problem ber words or Arabic numeral symbols. How- dependencies, a subcategory of so-called
has been formalized by Gold (100) and more ever, these studies reveal striking differences “finite-state grammars.” Despite their attrac-
recently and rigorously explored by Nowak in how animals and human children acquire tive simplicity, such rule systems are inade-
and colleagues (72–75). No known “general the integer list, and provide further evidence quate to capture any human language. Natu-
learning mechanism” can acquire a natural that animals lack the capacity to create open- ral languages go beyond purely local struc-
language solely on the basis of positive or ended generative systems. ture by including a capacity for recursive
negative evidence, and the prospects for find- Boysen and Matsuzawa have trained embedding of phrases within phrases, which
ing any such domain-independent device chimpanzees to map the number of objects can lead to statistical regularities that are
seem rather dim. The difficulty of this prob- onto a single Arabic numeral, to correctly separated by an arbitrary number of words or
lem leads to the hypothesis that whatever order such numerals in either an ascending or phrases. Such long-distance, hierarchical re-
system is responsible must be biased or con- descending list, and to indicate the sums of lationships are found in all natural languages
strained in certain ways. Such constraints two numerals (104 –106). For example, Boy- for which, at a minimum, a “phrase-structure
have historically been termed “innate dispo- sen shows that a chimpanzee seeing two or- grammar” is necessary. It is a foundational
sitions,” with those underlying language re- anges placed in one box, and another two observation of modern generative linguistics
ferred to as “universal grammar.” Although oranges placed in a second box, will pick the that, to capture a natural language, a grammar
these particular terms have been forcibly re- correct sum of four out of a lineup of three must include such capabilities (Fig. 5).
jected by many researchers, and the nature of cards, each with a different Arabic numeral. Recent studies suggest that the capacity to
the particular constraints on human (or ani- The chimpanzees’ performance might sug- compute transitional probabilities—an exam-
mal) learning mechanisms is currently unre- gest that their representation of number is like ple of a rule at the lowest level of the hierar-
solved, the existence of some such con- ours. Closer inspection of how these chim- chy—might be available to human infants
straints cannot be seriously doubted. On the panzees acquired such competences, howev- and provide a mechanism for segmenting
other hand, other constraints in animals must er, indicates that the format and content of words from a continuous acoustic stream
have been overcome at some point in human their number representations differ funda- (111–113). Specifically, after familiarization
evolution to account for our ability to acquire mentally from those of human children. In to a continuous sequence of consonant-vowel
the unlimited class of generative systems that particular, these chimpanzees required thou- (CV) syllables, where particular trigrams
includes all natural languages. The nature of sands of training trials, and often years, to (three CVs in sequence, considered to be
these latter constraints has recently become acquire the integer list up to nine, with no “words” in this context) have a high proba-
the target of empirical work. We focus here evidence of the kind of “aha” experience that bility of appearing within the corpus, infants
on the nature of number representation and all human children of approximately 3.5 are readily able to discriminate these tri-
rule learning in nonhuman animals and hu- years acquire (107). A human child who has grams from others that are uncommon.
man infants, both of which can be investigat- acquired the numbers 1, 2, and 3 (and some- Although this ability may provide a mech-
ed independently of communication and pro- times 4) goes on to acquire all the others; he anism for word segmentation, it is appar-
vide hints as to the nature of the constraints or she grasps the idea that the integer list is ently not a mechanism that evolved unique-
on FLN. constructed on the basis of the successor ly in humans or for language: The same
More than 50 years of research using clas- function. For the chimpanzees, in contrast, computation is spontaneously available to
sical training studies demonstrates that ani- each number on the integer list required the human infants for visual sequences and
mals can represent number, with careful con- same amount of time to learn. In essence, tonal melodies (113), as well as to nonhu-
trols for various important confounds (80). In although the chimpanzees’ understanding of man primates (cotton-top tamarins) tested
the typical experiment, a rat or pigeon is Arabic numerals is impressive, it parallels with the same methods and stimuli (114 ).
trained to press a lever x number of times to their understanding of other symbols and Similarly, in the same way that human
obtain a food reward. Results show that ani- their referential properties: The system appar- infants appear capable of computing alge-
mals can hit the target number to within a ently never takes on the open-ended genera- braic rules that operate over particular CV
closely matched mean, with a standard devi- tive property of human language. This limi- sequences (115), so too can cotton-top
ation that increases with magnitude: As the tation may, however, reveal an interesting tamarins (116 ), again demonstrating that
target number increases, so does variation quirk of the child’s learning environment and the capacity to discover abstract rules at a
www.sciencemag.org SCIENCE VOL 298 22 NOVEMBER 2002 1577You can also read
