Precis of: Foundations of Language: Brain, Meaning, Grammar, Evolution
Oxford and New York: Oxford
University Press, 2002 (506 pp.)
Ray Jackendoff
Program in Linguistics, MS 013
Brandeis University
Waltham, MA 02454 USA
jackendoff@brandeis.edu
tel: 617-484-5394
fax :617-484-0164
Short abstract:
The goal of this study to reintegrate the theory of generative grammar into the
cognitive sciences. Generative grammar was correct to focus on the child's
acquisition of language as its central problem, leading to the hypothesis of an
innate Universal Grammar. However,
generative grammar was mistaken to assume that the syntactic component is the
sole course of combinatoriality, and that everything else is
"interpretive." The proper approach is a parallel architecture, in
which phonology, syntax, and semantics are autonomous generative systems,
linked by interface components. The
parallel architecture leads to an integration within linguistics, and to a far
better integration with the rest of cognitive neuroscience. It fits naturally into the larger
architecture of the mind/brain and permits a properly mentalistic theory of
semantics. It leads to a view of
linguistic performance in which the rules of grammar are directly involved in
processing. Finally, it leads to a
natural account of the incremental evolution of the language capacity.
1.
Introduction
In the 1960s, when I became a graduate student in linguistics,
generative grammar was the hot new topic. Everyone from philosophers to psychologists to anthropologists to
educators to literary theorists was reading about transformational
grammar. But by the late 1970s, the
bloom was off the rose, although most linguists didn’t realize it; and by the
1990s, linguistics was arguably far on the periphery of the action in cognitive
science. To some extent, of course,
such a decline in fortune was simply a matter of fashion and the arrival of new
methodologies such as connectionism and brain imaging. However, there are deeper reasons for
linguistics’ loss of prestige, some historical and some scientific.
The basic questions I want to take up here, then, are:
What was right about generative grammar in the
1960s, such that it held out such promise?
What was wrong about it, such that it didn’t
fulfill its promise?
How can we fix it, so as to restore its value to the
other cognitive sciences?
The goal is to integrate linguistics with the other cognitive sciences,
not to eliminate the insights achieved by any of them. To understand language and the brain, we
need all the tools we can get. But
everyone will have to give a little in order for the pieces to fit together
properly.
The position developed in Foundations of Language is that the
overall program of generative grammar was correct, as was the way this program
was intended to fit in with psychology and biology. However, a basic technical mistake at the heart of the formal
implementation, concerning the overall role of syntax in the grammar, led to
the theory being unable to make the proper connections both within linguistic
theory and with neighboring fields. Foundations
of Language develops an alternative, the parallel architecture,
which offers far richer opportunities for integration of the field. In order to understand the motivation for
the parallel architecture, it is necessary to go through some history.
2. Three founding themes of generative grammar
The remarkable first chapter of Noam Chomsky’s Aspects of the Theory
of Syntax (1965) set the agenda for everything that has happened in
generative linguistics since. Three
theoretical pillars support the enterprise:
mentalism, combinatoriality, and acquisition.
Mentalism. Before Aspects, the predominant view
among linguists – if it was even discussed – was that language is something
that exists either as an abstraction, or in texts, or in some sense “in the
community” (the latter being the influential view of Saussure (1915), for
example). Chomsky urged the view that
the appropriate object of study is the linguistic system in the mind/brain of
the individual speaker. According to
this stance, a community has a common language by virtue of all speakers in the
community having essentially the same linguistic system in their minds/brains.1
The term most often used for this linguistic system is “knowledge”,
perhaps an unfortunate choice. However,
within the theoretical discourse of the time, the alternative was thinking of
language as an ability, a “knowing how” in the sense of Ryle (1949), which
carried overtones of behaviorism and stimulus-response learning, a sense from
which Chomsky with good reason wished to distance himself. It must be stressed, though, that whatever
term is used, the linguistic system in a speaker’s mind/brain is deeply
unconscious and largely unavailable to introspection, in the same way that our
processing of visual signals is deeply unconscious. Thus language is a kind of mind/brain property hard to associate
with the term “knowledge”, which commonly implies accessibility to
introspection. Foundations of
Language compromises with tradition by systematically using the term f-knowledge
(‘functional knowledge’) to describe whatever is in speakers’ heads that
enables them to speak and understand their native language(s).
There still are linguists, especially those edging off toward semiotics
and hermeneutics, who reject the mentalist stance and assert that the only
sensible way to study language is in terms of the communication between
individuals (a random example is Dufva and Lähteenmäki 1996). But on the whole, the mentalistic outlook of
generative grammar has continued to be hugely influential throughout
linguistics and cognitive neuroscience.
More controversial has been an important distinction made in Aspects
between the study of competence – a speaker’s f-knowledge of language –
and performance, the actual
processes (viewed computationally or neurally) taking place in the mind/brain
that put this f-knowledge to use in speaking and understanding sentences. I think the original impulse behind the
distinction was methodological convenience.
A competence theory permits linguists to do what they have always done,
namely study phenomena like Bulgarian case marking and Turkish vowel harmony,
without worrying too much about how the brain actually processes them. Unfortunately, in response to criticism
from many different quarters (especially in response to the collapse of the
derivational theory of complexity as detailed in e.g. Fodor, Bever, and Garrett
19xx), linguists have tended to harden the distinction into a firewall: competence theories were taken to be immune
to evidence from performance. And so
began a gulf between linguistics and
the rest of cognitive science that has persisted until the present.
Foundations
does not abandon the competence-performance distinction, but does return it to
its original status as a methodological rather than ideological
distinction. Although the innovations
in Foundations are largely in the realm of competence theory, one of
their important consequences is that there is a far closer connection to
theories of processing, as well as the possibility of a two-way dialogue
between competence and performance theories.
We return to this issue in section 9.3.
Combinatoriality: The earliest published work
in generative grammar, Chomsky’s Syntactic Structures (1957), began with
the observation that a language contains an arbitrarily large number of
sentences. Therefore, in addition to
the finite list of words, a characterization of a language must contain a set
of rules (or a grammar) that collectively describes or
“generates” the sentences of the language.
Syntactic Structures showed that the rules of natural language
cannot be characterized in terms of a finite-state Markov process, nor in terms
of a context-free phrase structure grammar.
Chomsky proposed that the appropriate form for the rules of a natural
language is a context-free phrase structure grammar supplemented by
transformational rules. Not all
subsequent traditions of generative grammar (e.g. Head-Driven Phrase Structure
Grammar (Pollard and Sag 1994) and Lexical-Functional Grammar (Bresnan 1982,
2001)) have maintained the device of transformational rules; but they all
contain machinery designed to overcome the shortcomings of context-free
grammars pointed out in 1957.2
Transferred into the mentalistic framework of 1965, the consequence of
combinatoriality is that speakers of the language must have rules of language
(or mental grammars) in their heads as part of their f-knowledge. Again there is a certain amount of
controversy arising from the term “rules”.
Rules of grammar in the sense of generative grammar are not like any of
the sorts of rules or laws in ordinary life:
rules of etiquette, rules of chess, traffic laws, or laws of
physics. They are unconscious
principles that play a role in the production and understanding of
sentences. Again, to ward off improper
analogies, Foundations uses the term f-rules for whatever the
combinatorial principles in the head may be.
Generative linguistics leaves open how directly the f-rules are involved in processing, but, as suggested
above, the unfortunate tendency among linguists has been not to care. The theory in Foundations, though,
makes it possible to regard the rules as playing a direct role in processing;
see again section 9.3.
An important reason for the spectacular reception of early generative
grammar was that it went beyond merely claiming that language needs rules: it offered rigorous formal techniques for
characterizing the rules, based on approaches to the foundations of mathematics
and computability developed earlier in the century. The technology suddenly made it possible to say lots of
interesting things about language and ask lots of interesting questions.
For the first time ever it was possible to provide detailed descriptions
of the syntax of natural languages (not only English but German, French,
Turkish, Mohawk, Hidatsa, and Japanese were studied early on). In addition, generative phonology took off
rapidly, adapting elements of Prague School phonology of the 1930s to the new
techniques. With Chomsky and Halle’s
1968 Sound Pattern of English as its flagship, generative phonology
quickly supplanted the phonological theory of the American structuralist
tradition.
Acquisition: Mentalism and combinatoriality together lead
to the crucial question: How do
children get the f-rules into their heads?
Given that the f-rules are unconscious, parents and peers cannot verbalize
them; and even if they could, children would not understand, since they don’t
know language yet. The best the
environment can do for a language learner is provide examples of the language
in a context. From there on it is up to
the language learner to construct the principles on his or her own –
unconsciously of course.
Chomsky asked the prescient question:
what does the child have to “(f-)know in advance” in order to accomplish
this feat? He phrased the problem in
terms of the “poverty of the stimulus”:
many different generalizations are consistent with the data presented to
the child, but the child somehow comes up with the “right” one, i.e. the one
that puts him or her in tune with the generalizations of the language
community. I like to put the problem a
bit more starkly: The whole community
of linguists, working together for decades with all sorts of crosslinguistic
and psycholinguistic data unavailable to children, has still been unable to
come up with a complete characterization of the grammar of a single natural
language. Yet every child does it by
the age of ten or so. Children don’t
have to make the choices we do: for
instance they don’t have to decide whether the “right” choice of grammar is in
the style of transformational grammar, the Minimalist Program, Optimality
Theory, Role and Reference Grammar, Tree-Adjoining Grammar, Cognitive Grammar,
connectionist networks, or some as yet unarticulated alternative. They already f-know it in advance.
One of the goals of linguistic theory, then, is to solve this “Paradox
of Language Acquisition” by discovering what aspects of linguistic f-knowledge
are not learned, but rather form the basis for the child’s
learning. The standard term for the
unlearned component is Universal Grammar or UG, a term that again
perhaps carries too much unwanted baggage.
In particular, UG should not be confused with universals of
language: it is rather what shapes the
acquisition of language. I prefer to
think of it as a toolkit for constructing language, out of which the child (or
better, the child’s brain) f-selects tools appropriate to the job at hand. If the language in the environment happens
to have a case system (like German), UG will help shape the child’s acquisition
of case; if it has a tone system (like Mandarin), UG will help shape the
child’s acquisition of tone. But if the
language in the environment happens to be English, which lacks case and tone,
these parts of UG will simply be silent.
What then is the source of language universals? Some of them will indeed be determined by
UG, for instance the overall “architecture” of the grammatical system: the parts of the mental grammar and their
relations (of which much more below).
Other universals, especially what are often called “statistical” or
“implicational” universals, may be the result of biases imposed by UG. For instance, UG may say that if a language
has a case system, the simplest such systems are thus-and-so; these will be
widespread systems crosslinguistically; they will be acquired earlier by
children; and systems may tend to change toward them over historical time. Other universals may be a consequence of the
functional properties of any relatively efficient communication system: for instance, the most frequently used
signals tend to be short. UG doesn’t
have to say anything about these universals at all; they will come about
through the dynamics of language use in the community (a process which of
course is not very well understood).
If UG is not learned, how does the child acquire it? The only alternative is through the
structure of the brain, which is determined through a combination of genetic
inheritance and the biological processes resulting from expression of the
genes, the latter in turn determined by some combination of inherent structure
and environmental input. Here
contemporary science is pretty much at an impasse. We know little about how genes determine brain structure and
nothing about how the details of brain structure determine anything about
language structure, even aspects of language as simple as speech sounds. Filling out this part of the picture is a
long-term challenge for cognitive neuroscience. It is premature to reject the hypothesis of Universal Grammar, as
some have (e.g. Elman et al. 1996 and Deacon 1997), arguing that we don’t know
how genes could code for language acquisition.
After all, we don’t know how genes code for birdsong or sexual behavior
or sneezing either, but we don’t deny that there is a genetic basis behind
these.
There next arises the question of how much of UG is a human cognitive
specialization for language and how much is a consequence of more general
capacities. The question has often been
oversimplified to a binary decision between language being entirely special or
entirely general, with a strong bias inside generative linguistics towards the
former and outside generative linguistics towards the latter. The truth of the matter undoubtedly lies
somewhere in between. To be sure, many
people (including myself) would find it satisfying if a substantial part of
language acquisition were a consequence of general human cognitive factors; but
the possibility of some specialization overlaying the general factors must not
be discounted. My view is that we
cannot determine what is general and what is special until we have comparable
theories of other cognitive capacities, including other learned
cognitive capacities. To claim that
language is parasitic on, say, motor control, perhaps because both have
hierarchical and temporal structure (this seems to be the essence of
Corballis’s (1991) position) – but without stating a theory of the f-knowledge
involved in motor control – is to
coarsen the fabric of linguistic theory to the point of unrecognizability. The closest approach to a comparable theory
is the music theory of Lerdahl and Jackendoff 1983, which displays some
striking parallels and some striking differences with language.
Of course, if UG – the ability to learn language – is in part a human
cognitive specialization, it must be determined by some specifically human
genes, which in turn had to have come into existence sometime since the hominid
line separated from the other great apes.
One would therefore like to be able to tell some reasonable story about
how it could be shaped by natural selection or other evolutionary
processes. We return to this issue in
section 9.4.
This approach to the acquisition of language has given rise to a
flourishing tradition of developmental research (references far too numerous to
mention) and a small but persistent tradition in learnability theory (e.g.
Wexler and Culicover 1980, Baker and McCarthy 1981). And certainly, even if the jury is not yet in on the degree to
which language acquisition is a cognitive specialization, there have been all
manner of phenomena investigated that bear on the issue, for instance:
- the sensitive period for language learning and the
consequences for first and second language acquisition at a later age
(Lenneberg 1967, Curtiss 1977, Newport 1990, Flynn and O’Neill 1988, Klein and
Perdue 1997)
- the limited ability of apes to acquire even
rudimentary versions of human language, even with extensive training (Premack
1976, Seidenberg and Petitto 1978, Terrace 1979, Savage-Rumbaugh et al. 1998)
- the characteristic brain localization of language
functions, resulting in characteristic aphasias (Zurif 1990)
- the grammatical parallels between spoken and
signed languages and the parallels in acquisition and aphasia (Klima and
Bellugi 1979, Bellugi, Poizner, and Klima 1989, Fischer and Siple 1990)
- the existence of characteristic language deficits
associated with various genetic conditions (Bellugi et al. 1994, Gopnik 1999,
Clahsen and Almazan 1998)
- the creation of creole languages by communities of
pidgin-speaking children (Bickerton 1981, DeGraff 1999)
- most strikingly, the creation of a signed language
de novo by a newly assembled community of deaf children in Nicaragua (Kegl et
al. 1999)
My impression is that, while there are questions about
all of these cases, en masse they offer an overwhelming case for some degree of
genetic specialization for language learning in humans.
These three foundational issues of generative grammar
– mentalism, combinatoriality, and acquisition – have stood the test of time;
if anything they have become even more important over the years in the rest of
cognitive science. It is these three
issues that connect linguistics intimately with psychology, brain science, and
genetics. Much of the promise of
generative linguistics arose from this new and exciting potential for
scientific unification.
3. The broken
promise: Deep Structure would be the key to the mind
A fourth major point of Aspects, and the one
that seeped most deeply into the awareness of the wider public, concerned the
notion of Deep Structure. A basic claim
of the 1965 version of generative grammar was that in addition to the surface
form of sentences, i.e. the form we hear, there is another level of syntactic
structure, called Deep Structure, which expresses underlying syntactic
regularities of sentences. For
instance, a passive sentence like (1a)
has a Deep Structure in which the noun phrases are in the order of the
corresponding active (1b).
(1) a. The bear was chased by the lion.
b. The lion chased the bear.
Similarly, a question such as (2a) has a Deep
Structure closely resembling that of the corresponding declarative (2b).
(2) a. Which martini did Harry drink?
b. Harry drank that martini.
In the years preceding Aspects, the question
arose of how syntactic structure is connected to meaning. Following a hypothesis first proposed by
Katz and Postal (1964), Aspects made the striking claim that the
relevant level of syntax for determining meaning is Deep Structure.
In its weakest version, this claim was only that
regularities of meaning are most directly encoded in Deep Structure, and this
can be seen in (1) and (2). However,
the claim was sometimes taken to imply much more: that Deep Structure IS meaning, an interpretation that
Chomsky did not at first discourage.3 And this was the part of generative
linguistics that got everyone really excited.
For if the techniques of transformational grammar lead us to meaning, we
can uncover the nature of human thought.
Moreover, if Deep Structure is innate – being dictated by Universal
Grammar – then linguistic theory gives us unparalleled access to the essence of
human nature. No wonder everyone
wanted to learn linguistics.
What happened next was that a group of generative
linguists, notably George Lakoff, John Robert Ross, James McCawley, and Paul
Postal, pushed very hard on the idea that Deep Structure should directly encode
meaning. The outcome, the theory of
Generative Semantics (e.g. McCawley 1968, Postal 1970, Lakoff 1971), increased
the “abstractness” and complexity of Deep Structure, to the point that the
example Floyd broke the glass was famously posited to have eight
underlying clauses, each corresponding to some feature of the semantics. All the people who admired Aspects
for what it said about meaning loved Generative Semantics, and it swept the
country. But Chomsky himself reacted
negatively, and with the aid of his then-current students (full
disclosure: present author included),
argued vigorously against Generative Semantics. When the dust of the ensuing“Linguistics Wars” cleared around
1973 (Newmeyer 1980, Harris 1993, Huck and Goldsmith 1995), Chomsky had won –
but with a twist: he no longer claimed
that Deep Structure was the sole level that determines meaning (Chomsky
1972). Then, having won the battle, he
turned his attention, not to meaning, but to relatively technical constraints
on movement transformations (e.g. Chomsky 1973, 1977).
The reaction in the larger community was shock: for one thing, at the fact that the
linguists had behaved so badly; but more substantively, at the sense that there
had been a “bait and switch.” Chomsky
had promised Meaning with a capital M and then had withdrawn the
offer. Many researchers, both inside
and outside linguistics, turned away from generative grammar with disgust,
rejecting not only Deep Structure but mentalism, innateness, and sometimes even
combinatoriality. And when, later in
the 1970s, Chomsky started talking about meaning again, in terms of a syntactic
level of Logical Form (e.g. Chomsky 1981), it was too late: the damage had been done. From this point on, the increasingly
abstract technical apparatus of generative grammar was of no interest to more than
a tiny minority of cognitive scientists, much less the general public.
Meanwhile, various nonChomskyan traditions of
generative grammar developed, most notably Relational Grammar (Perlmutter
1983), Head-Driven Phrase Structure Grammar (Pollard and Sag 1987, 1994),
Lexical-Functional Grammar (Bresnan 1982, 2001), Formal Semantics (Partee 1976,
Heim and Kratzer 19xx), Optimality Theory (Prince and Smolensky 1993),
Construction Grammar (Fillmore et al. 1988, Goldberg 1995), and Cognitive
Grammar (Lakoff 1987, Langacker 1987, Talmy 2000). On the whole, these approaches to linguistics (with the possible
exception of Cognitive Grammar) have made even less contact with philosophy,
psychology, and neuroscience than the recent Chomskyan tradition. My impression is that many linguists have
simply returned to the traditional concerns of the field: describing languages,
with as little theoretical and cognitive baggage as possible. While this is perfectly fine – particularly
since issues of innateness don’t play too big a role when you’re trying to record
an endangered language before its speakers all die – the sense of excitement
and danger that comes from participating in the integration of fields has
become attenuated.
4. The
scientific mistake: syntactocentrism
So much for pure intellectual history. We now turn to what I think was an important
mistake at the core of generative grammar, one that in retrospect lies behind
much of the alienation of linguistic theory from the cognitive sciences. Chomsky did demonstrate that language
requires a generative system that makes possible an infinite variety of
sentences. However, he explicitly
assumed, without argument (1965: 16, 17, 75, 198), that generativity is
localized in the syntactic component of the grammar – the construction of
phrases from words – and that phonology
(the organization of speech sounds) and semantics (the organization of meaning)
are purely “interpretive”, that is, that their combinatorial properties are
derived strictly from the combinatoriality of syntax.
In 1965 this was a perfectly reasonable view. The important issue at that time was to show
that something in language is generative. Generative syntax had provided powerful new tools, which were
yielding copious and striking results.
At the time, it looked as though phonology could be treated as a sort of
low-level derivative of syntax: the syntax gets the words in the right order,
then phonology massages their pronunciation to adjust them to their local
environment. As for semantics,
virtually nothing was known: the only
things on the table were the rudimentary proposals of Katz and Fodor (1963) and
some promising work by people such as Bierwisch (1967, 1969) and Weinreich
(1966). So the state of the theory
offered no reason to question the assumption that all combinatorial complexity
arises from syntax.
Subsequent shifts in mainstream generative linguistics
stressed major differences in outlook.
But one thing that remained unchanged was the assumption that syntax is
the sole source of combinatoriality.
Figure 1 diagrams the architecture of components in three major stages
of Chomskyan syntactic theory: the Aspects theory, Principles and
Parameters (or Government-Binding) Theory (Chomsky 1981), and the Minimalist
Program (Chomsky 1995). The arrows
denote direction of derivation.
Figure 1. Architecture of Chomsky's theories over the years.
These shifts alter the components of syntax and their
relation to sound and meaning. What
remains constant throughout, though, is that (a) there is an initial stage of
derivation in which words or morphemes are combined into syntactic structures;
(b) these structures are then massaged by various syntactic operations; and (c)
certain syntactic structures are shipped off to phonology/phonetics to be
pronounced and other syntactic structures are shipped off to “semantic
interpretation” to be understood. In
short, syntax is the source of all linguistic organization.
I believe that this assumption of “syntactocentrism” –
which, I repeat, was never explicitly grounded – was an important mistake at the heart of the field.4 The correct approach is to regard linguistic
structure to be the product of a number of parallel but interacting generative
capacities – at the very least, one each for phonology, syntax, and
semantics. As we will see, elements of
such a “parallel architecture” have been implicit in practice in the field for
years. What is novel in the present
work is bringing these practices out into the open, stating them as a
foundational principle of linguistic organization, and exploring the
large-scale consequences.
5. Phonology
as an exemplar of the parallel architecture
An unnoticed crack in the assumption of
syntactocentrism appeared in the middle to late 1970s, when the theory of
phonology underwent a major seachange.
Before then, the sound system of language had been regarded essentially
as a sequence of speech sounds. Any
further structure, such as the division into words, was thought of as simply
inherited from syntax. However, beginning with work such as Goldsmith (1979)
and Liberman and Prince (1977), phonology rapidly came to be thought of as
having its own autonomous structure, in fact multiple structures or tiers. Figure 2 provides a sample, the structure of
the phrase the big apple. The
phonological segments appear at the bottom, as terminal elements of the
syllabic tree.
Figure 2. Phonological structure of the big apple.
There are several innovations here. First, syllabic structure is seen as
hierarchically organized. At the center
of the syllable (notated as σ) is a syllabic nucleus (notated N),
which is usually a vowel but sometimes a syllabic consonant such as the l
in apple. The material following
the nucleus is the syllabic coda (notated C); this groups with the
nucleus to form the rhyme (notated R), the part involved in
rhyming. In turn, the rhyme groups with
the syllabic onset (notated O) to form the entire syllable. Syllables are grouped together into larger
units such as feet and phonological words (here, the bracketing subscripted Wd). Notice that in Figure 1, the word the
does not constitute a phonological word on its own; it is attached (or
cliticized) onto the word big.
Finally, phonological words group into larger units such as phonological
phrases. Languages differ in their
repertoire of admissible nuclei, onsets, and codas, but the basic hierarchical
organization and the principles by which strings of segments are divided into syllables
are universal. (It should also be
mentioned that signed languages have parallel syllabic organization, except
that the syllables are built out of manual rather than vocal constituents
(Klima and Bellugi 1979, Fischer and Siple 1990).)
These hierarchical structures are not built out of
syntactic primitives such as nouns, verbs, and determiners; their units are
intrinsically phonological. In
addition, the structures, though hierarchical, are not recursive.5 Thus the principles governing these structures
are not derivable from syntactic structures; they are an autonomous system of
generative rules.
Next consider the metrical grid in Figure 2. Its units are beats, notated as
columns of xs. A column with
only one x is a weak beat, and more xs in a column indicate a
relatively stronger beat. Each beat is
associated with a syllable; the strength of a beat indicates the relative
stress on that syllable, so that for example in Figure 2 the first syllable of apple
receives maximum stress. The basic principles
of metrical grids are in part autonomous of language: they also appear, for instance, in music (Lerdahl and Jackendoff
1983), where they are associated with notes instead of syllables. Metrical grids place a high priority on
rhythmicity: an optimum grid presents an alternation of strong and weak beats,
as is found in music and in much poetry.
On the other hand, the structure of syllables exerts an influence on the
associated metrical grid: syllables
with heavy rhymes (i.e. containing a coda or a long vowel) “want” to be
associated with relatively heavy stress.
The stress rules of a language concern the way syllabic structure comes
to be associated with a metrical grid; languages differ in ways that are now
quite well understood (e.g. Kager 1995, Halle and Idsardi 1995).
Again, metrical grids are built of nonsyntactic
units. As they are to some degree
independent of syllabic structure, they turn out to be a further autonomous
“tier” of phonological structure.
At a larger scale of phonological organization we find
prosodic units over which intonation contours are defined. These are comparable in size to syntactic
phrases but do not coincide with them.
Here are two examples.
(3) Syntactic
bracketing:
[Sesame
Street] [is [a production [of [the Children’s Television Workshop]]]]
Prosodic
bracketing (two pronunciations):
a. [Sesame Street is a production of] [the
Children’s Television Workshop]
b. [Sesame Street] [is a production] [of the
Children’s Television Workshop]
(4) Syntactic
bracketing
[This]
[is [the cat [that chased [the rat [that ate [the cheese]]]]]]
Prosodic
bracketing:
[This is
the cat] [that chased the rat] [that ate the cheese]
The two pronunciations of (3) are both acceptable, and
other prosodic bracketings are also possible.
However, the choice of prosodic bracketing is not entirely free, since
for instance [Sesame] [Street is a production of the] [Children’s Television
Workshop] is an impossible phrasing.
Now notice that the first constituent of (3a) and the second constituent
of (3b) do not correspond to any syntactic constituent. We would be hard pressed to know what
syntactic label to give to Sesame Street is a production of. But as an intonational constituent it
is perfectly fine. Similarly in (4),
the syntax is relentlessly right-embedded, but the prosody is flat and
perfectly balanced into three parts.
Again, the first two constituents of the prosody do not correspond to
syntactic constituents of the sentence.
The proper way to deal with this lack of correspondence
is to posit a phonological category of Intonational Phrase, which plays a role
in the assignment of intonation contours and the distribution of stress
(Beckman and Pierrehumbert 1986, Ladd 1996).
Intonation Phrases are to some degree correlated with syntax; their
boundaries tend to be at the beginning of major syntactic constituents; but
their ends do not necessarily correlate with the ends of the
corresponding syntactic constituents.
At the same time, intonational phrases have their own autonomous
constraints, in particular a strong preference for rhythmicity and parallelism
(as evinced in (2) for example), and a preference for saving the longest
prosodic constituent for the end of the sentence.6
Another example of mismatch between syntax and
phonology comes from contractions such as I’m and Lisa’s (as in Lisa’s
a doctor). These are clearly
phonological words, but what is their syntactic category? It is implausible to see them either as noun
phrases that incidentally contain a verb or to see them as verbs that
incidentally contain a noun. Keeping
phonological and syntactic structure separate allows us to say the natural
thing: they are phonological words that
correspond to two separate syntactic constituents.
(5) Syntactic
structure: [NP I] [V
(a)m] [NP Lisa] [V
(i)s]
Phonological
structure: [Wd
I’m] [Wd Lisa’s]
Since every different sentence of the language has a
different phonological structure, and since phonological structures cannot be
derived from syntax, the usual arguments for combinatoriality lead us to the
conclusion that phonological structure is generative. However, in addition to the generative principles that describe
these structures, it is necessary to introduce a new kind of principle into the
grammar, what might be called “correspondence rules” or “interface rules.” These rules (I revert to the standard term
“rules” rather than being obsessive about “f-rules”) regulate the way the
independent structures correspond with each other. For instance, the relation between syllable weight and metrical
weight is regulated by an interface rule between syllabic and metrical
structure; the relation between syntactic and intonational constituents is
regulated by an interface rule between syntactic and prosodic structure.
An important property of interface rules is they don’t
“see” every aspect of the structures they are connecting. For instance, the rules that connect
syllabic content to metrical grids are totally insensitive to syllable onset:
universally, stress rules care only about what happens in the rhyme. Similarly, although the connection between
syntax and phonology “sees” certain syntactic boundaries, it is insensitive to
the depth of syntactic embedding, Moreover, syntactic structure is totally
insensitive to the segmental content of the words it is arranging (e.g. there
is no syntactic rule that applies only to words that begin with b). Thus interface rules implement not
isomorphisms between the structures they relate, but rather only partial
homomorphisms.
This is not to say that we should think of speakers as
thinking up phonological and syntactic structures independently in the hope
they can be matched up by the interfaces.
That would be the same sort of mistake as thinking that speakers start
with the symbol S and generate a syntactic tree, finally putting in
words so they know what the sentence is about.
At the moment we are not thinking in terms of production; rather we are
stating the principles (of “competence”) in terms of which sentences are
well-formed. We will get back to how
this is related to processing in section 9.3.
Now the main point of this section. This view of phonological structure,
developed in the late 1970s and almost immediately adopted as standard, is
deeply subversive of the syntactocentric assumption that all linguistic
combinatoriality originates in syntax.
According to this view, phonological structure is not just a passive
hand-me-down derived from low-level syntax: it has its own role in shaping the
totality of linguistic structure. But
at the time of these changes, no great commotion was made about this most
radical aspect of the new phonology.
Phonologists for the most part were happy to get on with exploring this
exciting way of doing things, and for them, the consequences for syntax didn’t
matter. Syntacticians, for their part,
simply found phonology irrelevant to
their concerns of constraining movement rules and the like, especially since
phonology had now developed its own arcane technical machinery. So neither subdiscipline really took notice;
and as the technologies diverged, the relation between syntax and phonology
became a no-man’s-land (or perhaps only a very-few-man’s-land). Tellingly, as far as I can determine, in all
of Chomsky’s frequent writings on the character of the human language capacity,
there is no reference at all to post-1975 phonology – much less to the
challenge that it presents to his overall syntactocentric view of language.
6. The
syntax-semantics interface
I have treated the developments in phonology first
because it is less controversial. But
in fact the same thing happened in semantics.
Over the course of the 1970s and 1980s, several radically different
approaches to semantics developed: within
linguistics, at least Formal Semantics (growing out of formal logic)(Partee
1976, Heim and Kratzer 19xx), Cognitive Grammar (Lakoff 1987, Langacker 1987,
Talmy 2000), and Conceptual Semantics (Jackendoff 1983, 1990, Pinker 1989,
Pustejovsky 1995), plus approaches within computational linguistics and
cognitive psychology. Whatever their
differences, all these approaches take meaning to be deeply combinatorial. None of them take the units of semantic
structure to be syntactic units such as NPs and VPs; rather, the units are
intrinsically semantic entities like objects, events, actions, properties, and
quantifiers.7 Therefore,
whichever semantic theory we choose, it is necessary to grant semantics an
independent generative organization, and it is necessary to include in the
theory of grammar an interface component that correlates semantic structures
with syntactic and phonological structures.
In other words, the relation of syntax to semantics is qualitatively parallel
to the relation of syntax to phonology.
However, apparently no one pointed out the challenge to syntactocentrism
– except the Cognitive Grammarians, who mostly went to the other extreme and
denied syntax any independent role, and who have been steadfastly
ignored by mainstream generative linguistics.
The organization of phonological structure into
semi-independent tiers finds a parallel in semantics as well. Linguistic meaning can be readily partialled
into two independent aspects. On one
hand there is what might be called “propositional structure”: who did what to whom and so on. For instance, in The bear chased the lion,
there is an event of chasing in which the bear is the chaser and the lion is
“chasee”. On the other hand, there is
also what is now called “information structure”: the partitioning of the message into old vs. new information,
topic vs. comment, presupposition vs. focus, and so forth. We can leave the propositional structure of
a sentence intact but change its information structure, by using stress (6a-c)
or various focusing constructions (6d-f):
(6) a. The BEAR chased the lion.
b. The bear chased the LION.
c. The bear CHASED the lion.
d. It was the bear that chased the lion.
e. What the bear did was chase the lion.
f. What happened to the lion was the bear chased
it.
Thus the propositional structure and the information
structure are orthogonal dimensions of meaning, and can profitably be regarded
as autonomous tiers. (Foundations
proposes a further split of propositional structure into descriptive and
referential tiers, an issue too complex for the present context.)
Like the interface between syntax and phonology, that
between syntax and semantics is not an isomorphism. Some aspects of syntax make no difference in semantics. For instance, the semantic structure of a
language is the same whether or not the syntax marks subject-verb agreement,
verb-object agreement, or nominative and accusative case. The semantic structure of a language does
not care whether the syntax calls for the verb to be after the subject (as in
English), at
the end of the clause (as in Japanese), or second in a main clause and final in
a subordinate clause (as in German). As
these aspects of syntax are not correlated with or derivable from semantics,
the interface component disregards them.
Similarly, some aspects of semantics have little if
any systematic effect in syntax. Here
are a few well-known examples.
-
In example (8a), the interpretation is that Jill jumped
multiple times. This aspect of
interpretation does not arise from any single word in the sentence, nor from
the syntactic structure. If we change
the verb to sleep, as in (8b), we don’t interpret the sentence as
implying multiple acts of sleeping. If
we change until to when, as in (8c), only a single jump is
entailed.
(8) a. Jill
jumped until the alarm went off.
b. Jill slept until the alarm went off.
c. Jill jumped when the alarm went off.
The standard account of this contrast (Talmy 2000,
Verkuyl 1993, Pustejovsky 1995, Jackendoff 1997) is that the meaning of until
is to set a temporal bound on an ongoing process. When the verb phrase already denotes an ongoing process, such as
sleeping, all is well. But when the
verb phrase denotes an action that has a natural temporal ending, such as
jumping, then its interpretation is “coerced” into repeated action – a
sort of ongoing process – which in turn can have a temporal bound set on it by until. For present purposes, the point is that the
sense of repetition arises from semantic combination, without any direct
syntactic reflex. (On the other hand,
there are languages such as American Sign Language that have a grammatical
marker of iteration; this will have to be used in the translation of (8a).)
-
In the examples in (9), the “understood” subject of the
sentence is not the entity normally denoted by the actual subject (Nunberg
1979).
(9) a. [One
waitress says to another]:
The ham sandwich wants another cup of
coffee.
[Interpretation: ‘the person who
ordered/is eating the ham sandwich...’]
b. Chomsky is on the top shelf next to
Plato.
[Interpretation: ‘the book by
Chomsky ...’]
Such cases of “reference transfer” contain no syntactic
reflex of the italicized parts of the interpretation. One might be tempted to dismiss these phenomena as “mere
pragmatics”, hence outside the grammatical system. But this proves impossible, because reference transfer can have
indirect grammatical effects. A clear
example involves imagining that Richard Nixon went to see the opera Nixon in
China (yes, a real opera!), and what happened was that:
(10) Nixon was astonished to see himself sing
a foolish duet with Pat.
The singer of the duet, of course, is the actor
playing Nixon; thus the interpretation of himself involves a
reference transfer. However, we cannot
felicitously say that what happened next was that:
(11) *(Up on stage,) Nixon was astonished to
see himself get up and walk out.
That is, a reflexive pronoun referring to the acted
character can have the real person as antecedent, but not vice versa
(Fauconnier 1985, Jackendoff 1992).
Since the use of reflexive pronouns is central to grammar, reference
transfer cannot be seen as “extragrammatical.”
-
One of the most controversial issues within generative
grammar has been the syntactic status of quantifier scope. Consider the two interpretations of (12).
(12) Everyone in this room knows at least two
languages.
a. ‘John knows English and French; Sue
knows Hebrew and Hausa; ....”
b. ‘... namely, Mandarin and Navajo.’
Should there be two different syntactic structures
associated with these two interpretations?
Chomsky 1957 said no; Chomsky 1981 said yes; Generative Semantics said
yes; I am inclined to say no (Jackendoff 1996, Foundations chapter
12). The problem with finding two
different syntactic structures is that it requires systematic and drastic
distortions of the syntactic tree that never show up in the surface syntax of
any language. The problem with having
only one syntactic structure is that it makes the syntax-semantics interface
more complex. The point to be made here
is that the scope of quantification may well be a further example of the
“dirtiness” of the interface between syntax and semantics; this continues to be
an important issue in linguistic theory.
In each of these cases, a syntactocentric theory is
forced to derive the semantic distinctions from syntactic distinctions. Hence it is forced into artificial solutions
such as empty syntactic structure and elaborate movement, which have no
independent motivation beyond providing grist for the semantics. On the other hand, if the semantics is
treated as independent from syntax but correlated with it, it is possible to
permit a less than perfect correlation; it is then an empirical issue to
determine how close the match is.
If we abandon syntactocentrism, it is logically
possible that there are aspects of semantics that have no impact on syntax but do
have an effect on phonology through a direct phonology-semantics
interface. Such a treatment is
attractive for the correlation between prosody and information structure. For instance, the differences among (6a-c)
do not show up in syntax at all – only in the stress and intonation in
phonology, and in the focus-presupposition relations in semantics. In a syntactocentric theory, one is forced
to generate these sentences with a dummy syntactic element [+Focus], which
serves only to correlate phonology and meaning and does not affect word order
or inflection. (Such was the approach
in Jackendoff (1972), for instance.)
But this element does no work in syntax per se; it is only present in
order to account for the correlation between phonology and semantics. By introducing a direct
phonology-to-semantics interface sensitive to this correlation, we can account
for it with minimal extra machinery; but of course this requires us to
abandon syntactocentrism.
7. The
outcome: Parallel architecture
The argument so far has been that theoretical thinking
in both phonology and semantics has proceeded in practice as though their
structures are due to independent generative capacities. What has attracted far less notice among
syntacticians, phonologists, and semanticists alike is that such an
organization logically requires the grammar to contain interface components
that correlate the independent structures.
Carrying this observation through the entire architecture of grammar, we
arrive at an overall picture like Figure 3: the grammar contains multiple sets
of formation rules (the “generative” components), each determining its own
characteristic type of structure, and the structures are linked or correlated
by interface components.
Figure 3. The parallel archetecture.
In the syntactocentric architecture, a sentence is
well-formed when its initial syntactic tree is well-formed and all the steps of
derivation from this to phonology and semantics are well-formed. In the parallel architecture, a sentence is
well-formed when all three of its structures – phonological, syntactic, and
semantic – are independently well-formed and a well-formed correspondence among
them has been established by the interfaces.
One of the primary interface rules between phonology
and syntax is that the linear order of units in phonology corresponds to the
linear order of the corresponding units in syntax. One of the primary interface rules between syntax and semantics
is that a syntactic head (such as a verb, noun, adjective, or preposition)
corresponds to a semantic function, that the syntactic arguments of the head
(subject, object, etc.) correspond to the arguments of the semantic
function. The consequence of these two
primary interface principles is that for the most part, syntax has the linear
order of phonology but the embedding structure of semantics.
An illustration of some of these properties of the
parallel architecture appears in Figure 4, the structure of the phrase the
cats. The three independent
structures are displayed side by side;8 the subscripting
indicates the connections established by the interfaces between the parts of
the three structures. For example, the
clitic pronounced xc is
coindexed with the determiner in the syntax and the with definiteness feature
in semantics. Notice that the lowest
nodes in the syntactic tree are syntactic features, not the customary notation the
cat-s. The reasons for this are
explained in the next section.
Figure 4. The structure of the cats in the parallel architecture.
The overall architecture laid out in Figure 3 provides
a model within which many different theories of grammar can be embedded and
compared. For instance, Figure 3 does
not dictate whether the syntactic formation rules are along the lines of
transformational grammar, the Minimalist Program, Head-Driven Phrase Structure
Grammar, or many other alternatives.
Moreover, the syntactocentric framework is a version
of Figure 3 in which the phonological and semantic formation rules are null, so
that everything in phonological and semantic structures is determined only by
their interfaces with syntax. The
framework favored by many in Cognitive Linguistics minimizes or even eliminates
the syntactic formation rules, so that syntax is determined entirely by
meaning.
The organization into parallel generative components
is not new here. In addition to the
innovations in phonology discussed in section 3, Lexical-Functional Grammar
divides syntax into two tiers, c-structure and f-structure; Autolexical Syntax
(Sadock 1991) has a different division into morphosyntax and phrasal syntax;
Role and Reference Grammar (Van Valin and LaPolla 1997) has, in addition to a
morphosyntax/phrasal syntax division, the propositional/ information tier
division in semantics, with interfaces going every which way among the
tiers. In other words, various elements
of this architecture are widely present in the literature. What is novel here is recognizing that this
organization runs through the entire grammar, from phonology through semantics
(and further, as we will see in section 9).9
It might well be argued that the standard
syntactocentric framework has served the field well for forty years. Why should anyone want to give it up? A reply might come in five parts. First, no one has ever argued for the
syntactocentric model. In Aspects,
it was explicitly only an assumption, which quickly hardened into dogma and
then became part of the unstated background.
By contrast, the parallel architecture now has been argued for,
in part based on well-established results in phonology and semantics, which
have never played a role in syntactocentric argumentation.
Second, an advocate might argue that the
syntactocentric model is a priori simpler:
why should we admit so many different components into the grammar? The reply would be that the choice among
theories must be determined by empirical adequacy as well as a priori
simplicity. If the parallel
architecture allows a more perspicuous account of, say, intonation contours or
the relation of focus to stress, these are arguments in its favor.
A third point concerns the relation of syntax and
semantics. Since syntax is now not
responsible for determining every semantic distinction, it is to some degree
liberated from semantics and can therefore be considerably simplified. However, some compensating complexity must
be introduced into the syntax-semantics interface, so it doesn’t disappear from
the grammar entirely. It now becomes an
empirical question how to parcel the complexity out, and this question can be
addressed; it is not just an issue of opinion or preference (see e.g.
Jackendoff 1992, Culicover 1999, Culicover and Jackendoff 1995, 1997,
1999). At the same time, syntax does
not go away entirely (as opponents of Chomsky would often like). The syntax of a language still has to say
where the verb goes, whether the verb agrees with the subject, how to form
relative clauses and questions, and so on.
The differences among languages in these respects are not predictable
from semantics, and children have to learn them.
A fourth point concerns the nature of Universal
Grammar. In the parallel architecture,
the issues of acquisition and innateness don’t go away, they are exactly the
same, namely, How does the child acquire the grammar of its native language on
the basis of environmental evidence?
However, as just suggested, the sorts of questions that most often arise
concern the balance of power among components.
We don’t find ourselves invariably asking: what do we have to add to syntax to account for such-and-such a
phenomenon? Rather, we find ourselves
asking: in which component does this
phenomenon belong? Is it a fact of
syntax, of semantics, or of the interfaces?
And to what extent is it realistic to attribute such a bias to the child
learning the language?
A final point concerns not linguistic structure itself
but its connection to the rest of the theory of the brain/mind. On the face of it (at least in my opinion),
one should favor approaches that permit theoretical integration. Section 9 will show four ways that the
parallel architecture invites such integration but the syntactocentric theory
does not.
8. The lexicon
and the words vs. rules controversy
Every theory of language has to take a word to be a
stored complex of phonological, syntactic, and semantic features or structure;
commonly the store of words is called the lexicon. .
However, theories differ in the role of the lexicon in the construction
of sentences. In all of the
syntactocentric architectures shown in Figure 1, words are inserted into
syntactic trees at the beginning of a syntactic derivation, at the point when
syntactic trees are being built and before they begin to be manipulated and fed
to phonology and semantics. Thus the
traditional notation for trees in Figure 5a is actually intended as an
abbreviation of Figure 5b, in which the lexical items are spelled out in
full. The consequence is that the
syntax is carrying around with it all the phonological and semantic features of
words, which are totally invisible to syntactic rules and are of use to the
grammar only when handed over and “interpreted” by the proper component.
Figure 5a. Traditional notion for the cat. b. What the traditional notation abbreviates.
The parallel architecture, by contrast, insists that each
kind of feature belongs only in its own structure. Thus the traditional syntactic notation in Figure 5 is formally
incoherent, because it has phonological and semantic features in a syntactic
structure. Thus it is formally
impossible to insert full lexical items into syntactic structure. How then do words get into linguistic
structures? The answer is that each of
the three structures making up a word is inserted into its proper structure,
and each of them carries with it an index that connects it to the others. So, for example, the word cat is
notated as Figure 6; its contribution to the larger structure in Figure 4
should be evident.
Figure 6. The structure of the word cat.
Thus a word is best regarded as a type of interface
rule that establishes partial correspondences among pieces of phonological,
syntactic, and semantic structure (each piece in turn conforming to the
formation rules of its own component).
In other words, the language does not consist of a lexicon plus
rules of grammar; rather, lexical items are among the rules of grammar –
very particular rules to be sure, but rules nonetheless.
This treatment of the lexicon offers an attractive
account for a number of previously troublesome phenomena. For instance, consider an idiom such as kick
the bucket. This can be treated as
a lexically listed VP that is coindexed with phonology in the normal way, but
which lacks indexes connecting the individual words to semantics: instead, the
VP as a whole is coindexed with the semantic structure DIE. As a consequence, the individual words kick,
the, and bucket do not contribute individually to meaning. This is precisely what an idiom is supposed
to be: a stored unit in which the words
do not have their normal meaning.
A sort of converse is found in irregular
morphology. Consider something like the
irregular plural feet. It has to
be listed syntactically as a plural noun, and the two syntactic parts are
coindexed in the normal way to semantics:
the word denotes multiple entities of the type FOOT. However, the syntactic parts are not
connected in normal fashion to phonology; rather the whole syntactic complex is
coindexed with the undifferentiated lump feet in phonology.
Notice by contrast how the regular plural is
coded in Figure 4. The regular plural consists
of a piece of meaning, namely plurality, plus a piece of syntax, namely an
affix attached to nouns, plus a piece of phonology, namely a suffix s or
z or cz, the
choice determined contextually. That
is, the regular plural has all the same parts as a word, and it determines a
connection between them. We can notate
this as a lexical item along the lines of Figure 7. (The italicized bits denote contextual features that determine
how this item is combined with its environment.) The contribution of this item to the overall structure in Figure
4 is entirely parallel to the contribution of the word cat.
Figure 7. The English regular plural as a lexical item.
This view of regular
morphology puts a new and unexpected spin on the by now hoary “words vs. rules”
controversy (e.g. Rumelhart and McClelland 1986, Elman et al. 1996, Pinker
1999). Traditionally, the issue is
taken to be:
- Everyone
agrees that irregular plural nouns like feet have to be listed in the
lexicon. Are regular plurals all listed
as well, or is there a separate rule for the regular cases that says “To
form the plural of a noun, add -z”?
And, therefore, when children
learn to form regular plurals, are they learning something qualitatively
different from learning the rough-and-ready generalizations among irregular plurals?
On the present view,
words are rules – interface rules that help connect phonological,
syntactic, and semantic structures. And
Figure 7, the “rule” for the regular English plural affix, is qualitatively no
different. Its contextual features are
qualitatively not unlike those of, say, transitive verbs. It combines with nouns the same way a
transitive verb combines with its object.
Thus the formation of regular plurals is an instance of ordinary
combinatoriality. In this approach, the
issue comes to be restated like this:
- Are
regular plurals all listed, or is there a separate lexical item that
encodes the regular affix, which combines with any singular noun to form a
plural noun? And, therefore, when
children learn to form regular plurals, are they learning this new lexical item
by extracting it as a regularity from the contexts in which it appears – in the
same way that they extract verbs from the phrasal contexts in which they
appear?
I submit that even to the most committed of
connectionists, this latter way of framing the question can hardly be
objectionable.
In fact, the advocates of rules, such as Pinker, have
not made the case nearly as strong as it can be. The connectionist argument has been to the effect: We can make a device that learns all English
past tenses without making use of a rule, and we can find evidence from
acquisition and processing that supports this account. The best version of the anti-connectionist
argument has been: Connectionist
modeling offers important innovations over standard models of language in
dealing with case-by-case learning and analogy for the irregular past
tenses. But -- you still need rule
learning to account for children’s acquisition of regular past tense, and we
can find evidence from acquisition and processing that supports this
account. The problem is that the debate
has often been framed as though only the past tense were at issue, while the
subtext behind the connectionist position is that if this can be learned
without rules, then it is a good bet that the rest of language can be too.
But not only the past tense is at stake. To deal with the whole of language, it is
necessary to account for the creative formation of things like verb phrases and
relative clauses – linguistic entities that cannot be listed in the
lexicon. On the present view, the way
that the regular past tense affix combines with verbs is exactly like the way
nouns combine with relative clauses and the way noun phrases combine with verbs
and the way subordinate clauses combine with main clauses – it is just another
case of free combinatoriality. In the
decade and a half since the original connectionist past tense model, there has
been no demonstration that the model scales up to acquisition of the full free
combinatoriality of language – the issue that grounds generative
linguistics.
At the same time, notice that within the parallel
architecture, the terms of the dispute become far less contentious. The regular past tense is no longer a
qualitatively different phenomenon from words:
words are a type of rule, and the posited regular past tense morpheme,
Figure 7, is in the relevant respects just like a word. (It differs only in that it is grammatically
smaller and it requires a word as its grammatical host.) So the issue is only whether there is such a
separate lexical item, not whether there are two wholly different kinds of
linguistic animal, namely words and rules. Thus in the end the fate of the past
tense doesn’t seem like such a big deal.
9. Four ways the
parallel architecture helps integrate linguistics with cognitive neuroscience
The parallel architecture may be an intriguing
technical alternative to the Chomskyan orthodoxy in linguistics, but is there
any reason why it should be of interest to anyone other than linguists?
The end of the previous section may have begun to offer some hints. This section will sketch out a little more fully
some ways in which the parallel architecture offers opportunities to unify
linguistics with the other cognitive sciences.
9.1. The place
of the parallel framework in the larger architecture of the mind/brain.
To sum up the larger picture: The parallel
architecture claims that language is organized into a number of
semi-independent combinatorial systems, each of which has its own organizing
principles. These systems are linked by
systems of interface principles.
Interface principles establish a correlation between pieces of structure
in two (or more) of the combinatorial systems.
Some interface principles deal with large-scale and general
correspondences such as the parallel between linear order in syntax and in
phonology. On the other hand, some of
them are extremely specialized, for instance individual words, idioms, and
regular affixes. The interface principles
as a whole do not implement an isomorphism between the structures they
connect. Rather they implement a
partial homomorphism, a “dirty” correspondence in which not all parts of the
structures in question are correlated, and in which many-to-many mappings are
altogether common.
This conception of the interfaces within language is
perfectly in tune with the way linguistic structures connect to the rest of the
mind. Consider how phonology interacts
with the auditory system in speech perception and with the motor system in
speech production. As is well known (to
the dismay of fifty years of computer scientists working on automated speech
recognition), the mapping between a frequency analysis of the speech signal and
the phonological structure of an utterance is frighteningly complex. In particular, some aspects of the speech
signal play no role in phonological structure and must be factored out, for
instance the individual timbre of the speaker’s voice, the speaker’s tone of
voice, and the speed of production, not to mention ambient noise. These aspects of the speech signal are put
to use for other cognitive purposes, but not for speech. Moreover, having factored all these things
out from the acoustic signal, still not every part of the phonological
structure is predictable from what is left:
most prominently, word boundaries are not present as pauses in the
signal. Thus the
auditory-to-phonological mapping has the same general characteristics as the
interfaces inside language: it
establishes a “dirty” correspondence between certain aspects of two disparate
mental structures.
Speech production has similar properties. Not every aspect of phonological structure
corresponds to an aspect of the motor control involved in operating the vocal
tract. In particular, word boundaries
do not correspond at all consistently to pauses in production. And not every aspect of motor control is
controlled by phonological structure.
For instance, one can talk intelligibly with a pipe in one’s mouth,
which hugely distorts the motor commands involved in speech without changing
the phonological structure a bit. And
of course the same muscles in the vocal tract are used for chewing, swallowing,
and so on. Without going into more
detail, it should be clear that again the same sort of interface is in play
here.
Next consider the visual system. Beyond the very early levels of vision,
there is little detailed theory of the f-knowledge involved in vision – the
necessary levels of representation and so on (I take Marr (1983) to have been
attempting to lay out such a theory, but the enterprise has been largely
abandoned since his death). On the
other hand, the neuroscience of vision reveals a qualitatively similar picture: numerous independent brain areas, each
specializing in a particular aspect of vision such as shape, motion, color, and
spatial relations, each interacting with certain others by dedicated pathways,
and no area where “it all comes together” to form a full representation of the
visual field. This has precisely the
flavor of the parallel architecture in linguistics, where the notion of a
“sentence” or “phrase” is distributed among several structures, communicating
with each other via dedicated interfaces, as shown in Figure 4.10
A syntactocentric architecture, by comparison, shows
no resemblance to the rest of the mind/brain.
A master “computational system” that generates syntactic structures,
which in turn determine phonological structures and meanings, simply has no
known parallel in the brain. Even the
connection of language to speech is markedly different from the connections
among the components inside of language.
9.2. The role
of semantics
Another important advantage of the parallel
architecture is the connection of semantics to the rest of the mind/brain. Foundations of Language (chapters 9
and 10) advocates that if generative grammar is to truly adopt the mentalist
stance, this stance must be applied to meaning as well. According to this stance, the basic function
of language is to convert thoughts into communicable form; the virtue of human
language over other natural communication systems is that it is so broad in the
range of messages it can convey. Each
of the indefinitely many sentences of a language conveys a different
thought. Since not all these thoughts
can be stored in a single head, it is necessary that thoughts be constructed
combinatorially. Therefore a goal for
semantic theory is to uncover the combinatorial system underlying human
concepts. Such a goal converges with important
trends in psychology and philosophy.
However, another influential strain in semantics (and
the predominant one in Anglo-American philosophy, dating back to Frege (1892)
and shared by people as different as David Lewis (1972), Hilary Putnam (1976),
John Searle (1980), and Jerry Fodor (1987)) takes it that semantics is the
study of the connection of language to the world. On this view, a proper semantics has to be
concerned above all with how the noise kFt is
connected with cats. How language users
make that connection is quite a different issue (and to many semanticists, not
of interest). There is no room here to
begin the critique of this view; chapters 9 and 10 of Foundations take
up the argument in detail. My overall
conclusion is that even if it is worthwhile undertaking such a “realist”
semantics, the enterprise of discovering how language users do it is also
worthwhile. I don’t care whether you
call the latter enterprise semantics or shmenatics or whatever: it’s this enterprise whose central issues
intercalate naturally with those of generative linguistics, cognitive
psychology, and neuroscience. Just to
be clear, I will call this enterprise conceptualist semantics.
Conceptualist semantics requires us to rethink the
traditional issue of reference, which takes as its starting point the
unshakeable intuition that the phrase my cat does indeed pick out an
individual in the world. In a mentalist
linguistic theory, the language user’s linguistic system connects the phonological string /maykFt/ to the concept of a feline animal, and to the
concept of this feline animal being possessed by the speaker of the
phrase. How then does the language user
get from there to the actual individual out there in the world? The answer in brief is that it isn’t just
language users who have to connect something in their head to a sense of
individuals in the world: any organism
with a visual system about like ours (e.g. babies and apes) has precisely the
same problem. The environment acting on
the visual system produces some set of activations in the brain, resulting in
the organism experiencing real objects out there. In other words, conceptualist semantics allows us to recognize
that the problem of reference is not a problem about language, it’s at
bottom a problem about perception and cognition which has to be
solved by psychology and neuroscience.
By contrast, conventional realist theories of reference divorce
reference from the mind and make no contact whatsoever with research on
perception.
In order for the system of meaning to be influenced by
perception, of course, there has to be an interface between conceptual/semantic
structure and the “upper end” of the perceptual systems, where “the world”
(i.e. the perceiver’s conceptualization of the physical world) is organized in
terms of stable three-dimensional objects that are located in space with
respect to the perceiver and each other.
This interface too can be shown to have the standard characteristics: it
is a partial homomorphism between the quasi-algebraic format in which
linguistic meanings are encoded and the quasi-geometric/topological format(s)
in which spatial understanding is encoded.
Thus at the semantic end of the language faculty, just as at the phonological
end, the relation between language and the rest of the mind is of the same
general character as the interfaces within the language faculty itself.
Studying the conceptual system as a combinatorial
system leads to the same questions about acquisition as studying syntax. How does the child learning language acquire
the meanings of all those thousands of words on the basis of experience, both
perceptual and linguistic? What
perceptual biases and innate structures does the child bring to the task of
interpreting the world? Here conceptualist
semantics makes contact with a rich literature on word and concept learning and
its innate bases (Bloom 2000, Carey 1985, Spelke et al. 1994, Baillargeon 1986,
Gleitman and Landau 1994, Keil 1989, to mention only a few parochial examples). Moreover, since humans doubtless share with
monkeys and apes at least the parts of the conceptual system dealing with
physical space and perhaps some of the parts dealing with social relations and
other minds, conceptualist semantics further makes contact with research on
primate cognition (Premack 1976, Tomasello 2000, Hauser 2000, Povinelli 2001,
Cheney and Seyfarth 1990, Köhler 1927).
Again, these are issues that conventional realist
semantics cannot address. Nor are they
particularly accessible to semantics studied in a syntactocentric linguistic
theory. For if the combinatorial
properties of semantics were completely attributable to the combinatorial
properties of syntax, then it would be impossible for nonlinguistic organisms
to have combinatorial thoughts. There
are of course important strains of philosophy that have embraced this view,
identifying the capability for thought with the capability for overt language
(Descartes comes to mind, for instance).
But I think contemporary cognitive neuroscience has outgrown such a
view, and linguistics ought to be able to follow suit gracefully.
9.3. The
relation of grammar to processing
A theory of linguistic competence is supposed to
simply define the permissible structures in the language, without saying how
those structures are produced in real time.
However, as pointed out in
section 1, a competence theory ought to lend itself to being embedded in a
theory of processing: we ought to be
able to say how the f-knowledge that constitutes the competence theory is
actually put to use.
There turns out to be an inherent structural reason
why competence has to be isolated from performance in the syntactocentric view
of language. If we flatten out and
simplify all the syntactocentric architectures in Figure 1, they all have a
logical directionality proceeding outward from syntax in the middle:
(13) Logical
directionality of syntactocentric architecture
sound Z phonology
Z syntax Y meaning
What I mean by logical directionality is that the
possible phonological structures and meanings cannot be determined without
first determining syntactic structures.
Syntacticians may insist that they are being “metaphorical” when they
talk about things happening “before” and “after” other things in a derivation;
but the logical dependence is there nevertheless. Now contrast this to the logical
directionality of language processing:
language perception goes consistently from left to right, and language
production from right to left.
(14) a. Logical directionality of language
perception
sound Y phonology Y syntax Y meaning
b. Logical directionality of language
production
sound Z phonology
Z syntax Z meaning
Hence there is no way that the logical directionality
in (13) can serve the purposes of both perception and production. Going from syntax to phonology in (13) seems
inherently like production – but only part of production; going from
syntax to semantics in (13) seems inherently like perception – but only part of
it.
The parallel architecture, by contrast, is inherently nondirectional. The “information flow” between sound and
meaning is through the sequence of interfaces, each of which is a system of
correlations between two structures, not a derivation of one from the other. The correlations can be used in either
direction (which is why they are drawn with double arrows in Figure 3). This makes it possible to think of speech
perception as a process where structures are activated first at the auditory
end of the chain of structures, “clamped” by the environmental input. The interfaces propagate activation
rightward through the chain, each interface principle creating a partial
resonance between the structures it connects.
Eventually the structured activation reaches semantic structure, at
which point it can interact with the hearer’s understanding of the context to
produce the understanding of the heard utterance. Similarly, in speech production, the speaker begins with a
thought to convey, i.e. meaning is “clamped” by the speaker’s communicative
intent. Then the interface principles
propagate activation leftward through the chain, eventually activating motor
control of the vocal tract and producing speech. Crucially, except for the auditory and vocal parts of the chain,
the very same structures and the very same interface principles are invoked in
perception and production, just in opposite directions.
There is no need in this system for all of one level
to be totally processed before activation of the next level sets in. Any activation of a level, no matter how
incomplete, if it can be detected by the next interface, will start to
propagate to the next level in the chain.
Thus processing can be thought of as “incremental” or “opportunistic”
rather than rigidly regulated. In
addition, since the interfaces are trying to achieve “resonance”, i.e. optimal
mapping between levels, there is ample room in the processing theory for
feedback in processing – semantics affecting syntactic processing in
perception, and vice versa in production.
A crucial tenet of this theory, though, is that the
rules of grammar are the only source of information flow in language processing. For example, knowledge of context cannot
directly affect phonological processing, because there are no interface rules
that directly relate contextual understanding to phonological structure. On the other hand, context can indirectly
affect phonological processing – via the interfaces linking them through
semantics and syntax. The prediction is
that such feedback will take effect some time after constraints directly from
phonology, because it has to go up the chain of interfaces and down again. On the whole such a prediction seems
consistent with the experimental literature (Cutler and Clifton 1999, Levelt
1989); Foundations works out
many details, in particular the relation of long-term memory to working memory
during language processing.
The role of the lexicon in the processing theory is
entirely parallel to that in the competence theory. Recall that words are little interface rules, providing partial
routes for mapping between sound and meaning.
Now consider the logic of language perception. The auditory system and the interface from audition to phonology
produce some string of speech sounds in the hearer’s head, and this activates a
call to the lexicon: “Do any of you
guys in there sound like this?”
And various items raise their hand, i.e. get activated. At this point the processor has no way of
knowing which of these items is semantically appropriate, because no contact
has yet been made with semantics.
However, each item over time activates a connection to potential
syntactic and semantic structures, which can be integrated with previous words
and with context to determine which candidate word makes most sense in context. This scenario corresponds precisely to the
results in lexical access experiments (Swinney 1979, Tanenhaus et al. 1979), in
which every possible sense of a given phonological string is activated at
first, later to be pruned down by semantic context.
A parallel story can be told for speech
production. The speaker has activated
some conceptual structure that s/he wishes to communicate. The first step is to call the lexicon: “Do any of you guys in there mean this?” And various items raise their hand. All the lexical retrieval and speech error
literature now comes to bear in showing us the flow of information from this
point to actual vocal production; for the most part it proves to correspond
nicely to the options made possible by the components of the parallel framework
(Levelt 1989, 1999).
It is significant that the parallel architecture
accords words a very active role in determining the structure of sentences, in
concurrence with evidence from the psycholinguistic literature. By contrast, the syntactocentric
architecture views words as much more passive:
they simply ride around at the bottom of syntactic trees, while the
derivational rules of syntax do all the interesting work. Thus again, in the area of the lexicon, the
syntactocentric framework again makes it hard to connect competence and
performance.
The conclusion here is that the parallel architecture
permits a far closer relation between competence and performance theories. The rules of the language, including the
words, are posited to be precisely what the processing system uses in
constructing mappings between sound and meaning. This opens the door for a two-way dialogue between linguistics
and psycholinguistics. Linguistics has
always dictated the structures that psycholinguistics should be
investigating. But now there is the
possibility that psycholinguistic experiments may help determine what component
of the grammar is responsible for a particular phenomenon. For instance, PiZango et al. (1999) have shown that aspectual coercion
(e.g. the sense of repetition in Jill jumped until the alarm went off,
example (8a) above) causes a processing load at a point in time that is
appropriate to semantic, not syntactic processing. This result conforms to the theoretical claim that aspectual
coercion is a matter of adjusting semantic well-formedness, not a matter of
syntactic deletion of an iterative morpheme.
In short, the idealization of a competence theory is not a rigid
abstraction, it is rather a convenient methodological move, to be bridged
freely when the occasion arises. (More
detail on the relation of rules to processing is offered in Foundations,
chapter 7.)
9.4. Evolution
of language
Let us return to a point from section 2. If Universal Grammar is a human cognitive
specialization, it has to be transmitted by genes that have emerged in the
course of our evolutionary divergence from the chimpanzees. Of course the actual evidence for the
evolution of the language faculty is practically nonexistent. There is some evidence about the evolution
of the human vocal tract (Fitch 2000), but the ability to make speech sounds is
only one element of language – and of course there are signed languages, which
don’t involve speech at all. In
addition, it has begun to look like many of the mechanisms for auditory
perception are already in place in other mammals (Hauser, Chomsky, and Fitch
2002). But the real issue is: how did the ability to systematically map
combinations of concepts into sequences of speech sounds and back again develop
in our species, and how did the ability to learn such systematic
combinatorial mappings develop?
In the absence of evidence, we would like at least to
be able to tell a plausible story about the emergence of Universal Grammar, an
important aspect of which is the overall architecture of language. In particular, we would not like to have to
explain language through miraculous emergence, given that (as argued by Pinker
and Bloom 1990) it has the hallmarks of being shaped by natural selection. Pinker and Bloom, however, do not offer any
concrete proposals as to how language evolved.
As is well known, Chomsky himself has been notably evasive on the issue
of the evolution of the language faculty, often seeming to cast aspersions on
the theory of natural selection (Newmeyer 1998 collects representative
quotes). Chomsky is correct that other
factors besides natural selection play a role in evolution, for instance
environmental pressure and the biochemistry of proteins. Nevertheless there is nothing in these other
factors that provides any helpful hints on what brought about the emergence of
language.
The logic of the syntactocentric architecture suggests
a reason why such evasion has been necessary.
The problem is in providing a route for incremental evolution, such that
some primitive version of the faculty could still be useful to the
organism. In the syntactocentric
architecture, everything depends on syntax.
Meaning cannot have evolved before syntax, because its structure is
totally dependent on the syntactic structure from which it is derived. For the same reason, phonological structure
cannot have evolved before syntax. Thus
the complexity of syntax had to evolve before the complexity of the other components. But what would confer an adaptive advantage
on a syntactic faculty that just generated meaningless and imperceptible
syntactic structures? And what would
enable children to acquire such syntactic structures if there were no
perceptible output to which they could attach it? We quickly see that, at this very crude level at least, the
syntactocentric theory is stuck: there
is no logical way to build it incrementally, such that the earlier stages are
useful.
The parallel architecture offers a better
alternative. The system of concepts
that language expresses is an independent generative component in the
mind/brain. Since it is taken to exist,
in some degree at least, in other primates as well, it also could have existed
in our ancestors, prior to language.
That is, our ancestors had interesting thoughts, but lacked any way to
say them: meaning therefore would be
the first generative component of language to emerge (a similar view is urged
by Hauser 2000).
Most speculation on language evolution goes on to say
that the earliest stage would have been the symbolic use of simple
vocalization, without grammatical organization. Such a stage is theoretically impossible in the syntactocentric
theory, since even single-word utterances have to arise from syntactic
structure. But such a stage is quite
natural in the parallel architecture:
it consists of stored associations of vocalizations and concepts, i.e. a
“paleo-lexicon.” Lexical items that can
serve on their own as utterances still exist in modern language, for instance hello,
oops, ouch, and gadzooks.
The provision for them in language might be viewed as an evolutionary
relic of this earliest stage.
Assuming that there would be an adaptive advantage to
a larger number of signals, a regimentation of vocalization along the lines of
phonological structure would be the next generative component of language to
emerge. Phonological organization in
effect digitizes vocalizations, making a large vocabulary reliably
discriminable and learnable. (Proto-)words
at this point would be simply duples of phonological and semantic structure,
without syntax.
A next innovation might be the provision of
concatenating words into larger utterances.
However, when words are concatenated, the issue arises of how the
meanings of words in a string are related to each other semantically. In a string like eat apple Fred, it
is pretty clear on pragmatic grounds that Fred is eating the apple and not the
reverse. But pragmatics can only go so
far: in chase lion bear, who is
the chaser? Something as elaborate as
English syntax is not entirely necessary to fix this. One can actually get considerable mileage from simple functional
principles of linear word order. For
example, the principle “Agent First” would tell us that the lion is chasing the
bear and not the reverse. Such a
principle is a straight phonology-to-semantics mapping, relating linear order
to semantic function. And principles
like this appear to be widespread in pidgin languages (Bickerton 1981) and the
grammars of speakers who have acquired their languages late in life, after the
sensitive period (Klein and Perdue 1997).
Finally, principles like Agent First have their
limitations too. One can imagine the
capacity for modern syntactic structure evolving last, as a way of
making more complex semantic relations among the words of an utterances more
precisely mappable to linear word order in phonology. That is, syntax comes along in evolution as a refinement, a
“supercharger” of a preexisting interface between phonology and semantics. This seems exactly appropriate to its
function within the parallel architecture.
In short, the parallel architecture cannot tell us
exactly how language evolved – I don’t think anything can ever tell us
that. But it does afford a far more
plausible hypothesis than the syntactocentric architecture (Foundations,
chapter 8 develops this story in considerably more detail). Thus the parallel architecture opens the
door for linguistics to participate far more fully in the mainstream of
evolutionary psychology, yet another desirable connection.
10.
Conclusions
Putting this all together, the parallel architecture
makes it possible both to internally integrate linguistic theory, establishing
the proper relation between phonology, syntax, semantics, and the lexicon, and
also to integrate linguistic theory more comprehensively with the brain and
with biology. In addition, by
liberating semantics from its syntactic shackles, the parallel architecture
makes it possible to develop a fully psychological theory of meaning and its
relation to perception. As observed
above, these points of connection were precisely what early generative grammar
promised but ultimately couldn’t bring off; I have tried to show here why
syntactocentrism was a major reason behind this disappointment.
Of course, to propose a new architecture only begins
the work. It opens major questions
about exactly what components the grammar requires and what interfaces connect
them. Vast numbers of phenomena have
been studied in the context of the traditional architecture; to what extent can
the analyses proposed there be duplicated or even improved upon? In particular, a thorough overhaul of
syntactic theory is necessary, in order to overcome decades of accretions
motivated solely by syntactocentric
assumptions (Culicover and Jackendoff (forthcoming) begin to undertake this
task). Perhaps the hardest part of all
this will be maintaining a sense of global integration, keeping the subdomains
of the field in closer touch than they have recently been.
But linguistics alone cannot sustain the weight of the
inquiry. We need all the help we can
get from every possible quarter. And in
return, one would hope that linguistic theory might be a more fruitful source of
evidence and puzzles for other fields.
Above all, my aspiration for Foundations is that it can help
encourage the necessary culture of collaboration.
References
Baillargeon, Renée (1986) Representing the existence and the location of hidden objects: Object permanence in 6- and 8-month-old
infants. Cognition 23, 21-41.
Baker, C. L., and John J. McCarthy (eds.) (1981) The logical problem of language
acquisition. Cambridge, MA: MIT Press.
Beckman, Mary, and Janet Pierrehumbert (1986)
Intonational structure in English and Japanese. Phonology 3, 255-309. Bellugi et al. 1994
Bellugi, Ursula, Howard Poizner, and Edward S. Klima
(1989) Language, modality, and the
brain. Trends in Neurosciences
12, 380-388.
Bickerton, Derek (1981) Roots of language.
Ann Arbor, MI: Karoma.
Bierwisch, Manfred (1967) Some semantic universals of German adjectivals. Foundations of Language 3, 1-36.
Bierwisch, Manfred (1969) On certain problems of semantic representation. Foundations of Language 5, 153-184.
Bloom, Paul (2000) How children learn the meanings
of words. Cambridge, Mass.: MIT
Press.
Bresnan, Joan W. (ed.) (1982) The mental
representation of grammatical relations.
Cambridge, MA: MIT Press.
Bresnan, Joan W. (2001) Lexical-functional syntax. Oxford: Blackwell.
Carey, Susan (1985) Conceptual Change in Childhood. MIT Press.
Cheney, Dorothy, and Robert Seyfarth (1990) How monkeys see the world. Chicago:
University of Chicago Press.
Chomsky, Noam (1957)
Syntactic structures. The
Hague: Mouton.
(1965) Aspects of the theory of syntax. Cambridge, MA: MIT Press.
(1966) Cartesian Linguistics. New York:
Harper & Row.
(1972) Studies on Semantics in Generative
Grammar. The Hague: Mouton.
(1973) Constraints on transformations. In S. Anderson and P. Kiparsky (eds.), A
Festschrift for Morris Halle, 232-286.
New York: Holt, Rinehart &
Winston.
(1977) On wh-movement. In P. Culicover, T. Wasow, and A. Akmajian (eds.), Formal
syntax, 71-132. New York: Academic Press.
(1981) Lectures on government and binding. Dordrecht: Foris.
(1995) The Minimalist Program. Cambridge, MA: MIT Press.
Chomsky, Noam, and Morris Halle (1968) The sound pattern of English. New York, Harper & Row.
Clahsen, Harald, and Mayella Almazan (1998) Syntax and morphology in Williams
syndrome. Cognition 68, 167-198.
Corballis, Michael C. (1991) The lopsided ape.
Oxford: Oxford University Press,
Culicover, Peter W. (1999) Syntactic nuts: Hard
cases in syntax. Oxford: Oxford
University Press.
Culicover, Peter, and Ray Jackendoff (1995) Something else for the binding
theory. Linguistic Inquiry 26,
249-275.
(1997) Semantic subordination despite syntactic
coordination. Linguistic Inquiry
28, 195-217.
(1999) The view from the periphery: The English correlative conditional. Linguistic Inquiry 30, 543-571.
(forthcoming)
Minimal syntax.
Curtiss, Susan (1977)
Genie: A linguistic study of
a modern-day "wild child".
New York, Academic Press.
Cutler, Anne, and Charles Clifton, Jr. (1999) Comprehending spoken language: A blueprint of the listener. In Colin M. Brown and Peter Hagoort (eds.), The
neurocognition of language, 123-166.
Oxford: Oxford University Press.
Deacon, Terrence W. (1997) The Symbolic Species.
New York: Norton.
Degraff, Michel (ed.) (1999) Language creation and language change: Creolization, diachrony, and development. Cambridge, MA: MIT Press.
Dufva, H., and M. Lähteenmäki (1996) But who killed Harry? A dialogical approach
to language and consciousness. Pragmatics
and Cognition 4, 35-53.
Elman, Jeffrey (1990)
Finding structure in time. Cognitive
Science 14, 179-211.
Elman, Jeffrey, Elizabeth Bates, Mark Johnson, Annette
Karmiloff-Smith, Domenico Parisi, and Kim Plunkett (1996) Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press.
Fauconnier, Gilles (1985) Mental spaces: Aspects
of meaning constrution in natural language. Cambridge, MA: MIT Press.
Fillmore, Charles, Paul Kay, and Mary Catherine
O'Connor (1988) Regularity and idiomaticity in grammatical constructions: The case of let alone. Language 64, 501-538.
Fischer, Susan D.,
and Patricia Siple (eds.) (1990) Theoretical issues in sign language
research 1. Chicago: University of
Chicago Press.
Fitch, W. Tecumseh (2000) The evolution of speech:
A comparative review. Trends
in Cognitive Sciences 4, 258-267.
Flynn, Suzanne, and Wayne O'Neill (eds.) (1988) Linguistic theory in second language
acquisition. Dordrecht: Reidel.
Fodor, Jerry (1987)
Psychosemantics: The problem
of meaning in the philosophy of mind.
Cambridge, MA: MIT Press.
Fodor, Jerry, Thomas Bever, and Merrill Garrett
(1974) The Psychology of Language. New York:
McGraw-Hill.
Frege, Gottlob (1892)
Über Sinn und Bedeutung. Zeitschrift
für Philosophie und philosophische Kritik 100, 25-50. English translation in Peter Geach and Max
Black (eds.), Translations from the philosophical writings of Gottlob Frege,
56-78. Oxford: Blackwell, 1952.
Gleitman, Lila R., and Barbara Landau (eds.)
(1994) The acquisition of the
lexicon. Cambridge, MA: MIT Press.
Goldberg, Adele (1995) Constructions: A Construction Grammar approach to argument
structure. Chicago: University of Chicago Press.
Goldsmith, John (1976) Autosegmental phonology.
Doctoral dissertion, MIT. New
York: Garland Press, 1979.
Gopnik, Myrna (1999)
Some evidence for impaired grammars.
In R. Jackendoff, P. Bloom, and K. Wynn (eds.), Language, Logic, and
Concepts, 263-283. Cambridge,
Mass.: MIT Press.
Halle, Morris, and William Idsardi (1995) Stress and metrical structure. In J. Goldsmith (ed.), Handbook of
theoretical phonology, 403-443.
Oxford: Blackwell.
Harris, Randy Allen (1993) The Linguistics Wars.
New York: Oxford University
Press.
Hauser, Marc D. (2000) Wild Minds: What
Animals Really Think. New
York: Henry Holt.
Hauser, Marc D., Noam Chomsky, and Tecumseh Fitch
(2002) Science
Heim, Irena, and Angelica Kratzer (199x)
Huck, Geoffrey, and John Goldsmith (1995) Ideology and linguistic theory. Chicago:
University of Chicago Press.
Jackendoff, Ray (1972) Semantic interpretation in generative grammar. Cambridge, MA: MIT Press.
(1983) Semantics and cognition. Cambridge, MA: MIT Press.
(1992) Mme. Tussaud meets the binding theory. Natural Language and Linguistic Theory
10.1, 1-31.
(1990) Semantic structures. Cambridge, MA: MIT Press.
(1996) The proper treatment of measuring out,
telicity, and possibly even quantification in English. Natural Language and Linguistic Theory
14, 305-354.
(1997) The architecture of the language faculty. Cambridge, MA: MIT Press.
Kager, René (1995)
The metrical theory of word stress.
In J. Goldsmith (ed.), Handbook of theoretical phonology,
367-402. Oxford: Blackwell.
Katz, Jerrold, and Jerry Fodor (1963) The structure of a semantic theory. Language 39, 170-210.
Katz, Jerrold, and Paul M. Postal (1964) An integrated theory of linguistic
descriptions. Cambridge, MA: MIT Press.
Kegl, Judy, Ann Senghas, and Marie Coppola (1999) Creations through contact: Sign language emergence and sign language
change in Nicaragua. In DeGraff 1999,
179-237.
Keil, Frank C. (1989)
Concepts, kinds, and cognitive development. Cambridge, MA: MIT Press.
Klein, Wolfgang, and Clive Perdue (1997)
The Basic Variety, or: Couldn't language be much simpler? Second Language Research 13, 301-347.
Klima, Edward S., and Ursula Bellugi (1979) The signs of language. Cambridge, MA: Harvard University Press.
Köhler, Wolfgang (1927) The mentality of apes, Routledge & Kegan Paul.
Ladd, D. Robert (1996) Intonational phonology.
Cambridge: Cambridge University Press.
Lakoff, George (1971)
On generative semantics. In D.
Steinberg and L. Jakobovits (eds.), Semantics: An interdisciplinary reader in philosophy, linguistics, and
psychology, 232-296. New York: Cambridge University Press.
Lakoff, George (1987) Women, fire, and dangerous
things. Chicago: University of Chicago Press.
Lamb, Sydney (1966) Outline of Stratificational
Grammar. Washington: Georgetown
University Press.
Langacker, Ronald (1987) Foundations of cognitive grammar, Vol. 1. Stanford:
Stanford University Press.
Lenneberg, Eric (1967) Biological foundations of language. New York:
John Wiley & Sons.
Lerdahl, Fred, and Ray Jackendoff (1983) A Generative Theory of Tonal Music. Cambridge, MA: MIT Press.
Levelt, W. J. M. (1989) Speaking.
Cambridge: MIT Press.
(1999) Producing spoken language: A blueprint of the speaker. In Colin M. Brown and Peter Hagoort (eds.), The
neurocognition of language, 83-122.
Oxford: Oxford University Press.
Liberman, Mark, and Alan Prince (1977) On stress and linguistic rhythm. Linguistic Inquiry 8, 249-336.
Marcus, Gary (2001)
The algebraic mind. Cambridge, MA: MIT Press.
Marr, David (1982)
Vision. San
Francisco: Freeman.
McCawley, James D. (1968) Lexical insertion in a transformational grammar without deep
structure. In B. Darden, C.-J. N.
Bailey, and A. Davison (eds.) Papers from the fourth meeting of the Chicago Linguistic
Society. Chicago: University of Chicago Department of
Linguistics.
Newmeyer, Frederick J. (1980) Linguistic theory in America: The first quarter-century of
transformational-generative grammar.
New York: Academic Press.
(1998) On the supposed 'counterfunctionality' of
Universal Grammar: some evolutionary
implications, in James Hurford, Michael Studdert-Kennedy, and Chris Knight
(eds.) Approaches to the evolution of language, 305-319. Cambridge:
Cambridge University Press.
Newport, Elissa (1990) Maturational constraints on language learning. Cognitive Science 14, 11-28.
Nunberg, Geoffrey (1979) The non-uniqueness of semantic solutions: Polysemy.
Linguistics and Philosophy 3, 143-184.
Partee, Barbara, ed. (1976) Montague Grammar.
New York, Academic Press.
Perlmutter, David M. (ed.) (1983) Studies in Relational Grammar 1. Chicago:
University of Chicago Press.
PiZango, Maria
Mercedes, Edgar Zurif, and Ray Jackendoff (1999) Real-time processing implications of enriched composition at the
syntax-semantics interface. Journal
of Psycholinguistic Research, 28, 395-414.
Pinker, Steven (1989)
Learnability and cognition:
The acquisition of argument structure. Cambridge, MA: MIT Press
(1999) Words and rules. New York: Basic Books.
Pinker, Steven, and Paul Bloom (1990) Natural language and natural selection, Behavioral
and Brain Sciences 13, 707-726
Pollard, Carl, and Ivan Sag (1987) Information-based syntax and semantics. Stanford: Center for the Study of Language
and Information.
(1994) Head-Driven Phrase Structure Grammar. Chicago: University of Chicago Press.
Postal, Paul M.
(1970) On the surface verb
"remind." Linguistic
Inquiry 1, 37-120.
Povinelli, Daniel (2002) Folk Physics for Apes. Oxford: Oxford University Press.
Premack, David (1976)
Intelligence in ape and man.
Hillsdale, NJ: Erlbaum.
Prince, Alan, and Paul Smolensky (1993). Optimality Theory: Constraint Interaction
in Generative Grammar. Rutgers
University Center for Cognitive Science, Piscataway, NJ.
Pustejovsky, James (1995) The generative lexicon.
Cambridge, MA: MIT Press.
Putnam, Hilary (1975)
The meaning of 'meaning.' In
Keith Gunderson (ed.), Language, mind, and knowledge, 131-193. Minneapolis: University of Minnesota Press.
Rumelhart, David, and James McClelland (1986) On learning the past tense of English
verbs. In J. McClelland, D. Rumelhart,
and the PDP Research Group, Parallel distributed processing, Vol. 2,
216-271. Cambridge, MA: MIT Press.
Ryle, Gilbert (1949)
The concept of mind.
Chicago: University of Chicago
Press.
Sadock, Jerrold (1991) Autolexical syntax.
Chicago: University of Chicago
Press.
Saussure, Ferdinand de (1915) Cours de linguistique générale,
edited by Charles Bally and Albert Sechehaye.
English translation: Course
in General Linguistics. New
York: Philosophical Library, 1959.
Savage-Rumbaugh, Sue, Stuart Shanker, and Talbot
Taylor (1998) Apes, language, and
the human mind. Oxford University
Press.
Searle, John (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3,
417-424.
Seidenberg, Mark S., and Laura Petitto (1978) Signing behavior in apes: A critical review. Cognition l7, 177-215.
Spelke, Elizabeth, Gary Katz, Susan Purcell, Sheryl
Ehrlich, and Karen Breinlinger (1994)
Early knowledge of object motion:
Continuity and inertia. Cognition
51, 131-176.
Swinney, David (1979) Lexical access during sentence
comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior 18, 645-659.
Talmy, Leonard (2000) Toward a Cognitive Semantics. Cambridge, Mass.: MIT Press.
Tanenhaus, M. K., M. J. Spivey-Knowlton, K. M.
Eberhard, and J. C. Sedivy (1995).
Integration of Visual and Linguistic Information in Spoken Language
Comprehension. Science 268,
1632-4.
Terrace, H. (1979) Nim. New York, Knopf.
Tomasello, Michael (ed.) (2000) Primate cognition (special
issue). Cognitive Science 24.3.
Van Valin, Robert, and Randy LaPolla (1997) Syntax:
Structure, meaning and function.
Cambridge: Cambridge Unviersity
Press.
Verkuyl, Henk (1993)
A theory of aspectuality: The
interaction between temporal and atemporal structure. Cambridge:
Cambridge University Press.
Weinreich, Uriel (1966) Explorations in semantic theory.
In T. Sebeok (ed.), Current trends in linguistics, vol. 3. The Hague:
Mouton. Reprinted in U.
Weinreich (1980), On semantics, 99-201.
Philadelphia: University of
Pennsylvania Press.
Wexler, Kenneth, and Peter Culicover (1980) Formal principles of language acquisition. Cambridge, MA: MIT Press.
Zurif, Edgar (1990)
Language and the brain. In Dan
Osherson and Howard Lasnik (eds.), An introduction to cognitive science,
Vol. 1: Language, 177-198. Cambridge, MA: MIT Press.
Acknowledgment
Foundations of Language was written during 1999-2000, while I was a Fellow at
the Wissenschaftskolleg zu Berlin, a year for which I am profoundly
grateful. Both the book and this
version of it were supported in part by NIH Grant 03660 to Brandeis
University. I also must thank Marc
Hauser and Dan Dennett for their help and encouragement on this version.
Notes