<< . .

( 19)

. . >>

in essence it is there independently of language. The bones of the
middle ear don™t change because of language. And there is some kind
of system of thought (conception, intention and so on) which is sort
of sitting there. That includes what were traditionally called “common
notions” or “innate ideas.” Perhaps also analysis in terms of what
is called “folk psychology,” interpreting people™s actions in terms of
belief and desire, recognizing things in the world and how they move,
and so on. Well, that™s presumably not entirely dependent on lan-
guage; probably, non-human primates have something like that, and
perhaps even the capacity of attributing minds to other organisms, a
question currently much debated. The language faculty has to interact
with those systems, otherwise it™s not usable at all. So, we may ask: is
it well designed for the interaction with those systems? Then you get a
different set of conditions. And in fact the only condition that emerges
clearly is that, given that the language is essentially an information
system, the information it stores must be accessible to those systems,
that™sthe only condition. We can ask whether language is well designed
to meet the condition of accessibility to the systems in which it is
embedded. Is the information it provides “legible” to those systems?

An interview on minimalism

It is like asking: is the liver accessible to the other systems with which
it interacts? If the liver produced something, not bile, but something
else that the rest of the body couldn™t make any use of, it wouldn™t be
any good; and that™s a different question than whether the liver is well
designed for life in a wine-drinking culture. A very different question.
ab & lr: An empirically non-vacuous definition of perfection
implies the identification of possible imperfections. Inflectional mor-
phology is often referred to as an apparent imperfection. For instance,
invented formal languages have a recursive syntax, capable of comput-
ing expressions over an unbounded domain, but nothing resembling
natural language morphology. What is the driving intuition here?
Morphology seems to be at the same time an imperfection and
a defining property of natural languages. How can these two aspects
be reconciled within a minimalist perspective?
nc: Morphology is a very striking imperfection; at least, it is
superficially an imperfection. If you were to design a system, you
wouldn™t put it in. It™s not the only one, though; no formal language,
for example, has a phonology or a pragmatics and things like dislo-
cation in the sense we all understand: expressions appear not where
you interpret them but somewhere else. All of these are imperfections,
in fact even the fact that there is more than one language is a kind of
imperfection. Why should that be? All of these are at least prima facie
imperfections, you would not put them into a system if you were trying
to make it work simply. A good guiding intuition about imperfection
is to compare natural languages with invented “languages,” invented
symbolic systems. When you see differences, you have a suspicion that
you are looking at something that is a prima facie imperfection. There
are differences at about every point. Formal languages, for example,
don™t have a designated syntax; they just have a set of well-formed
expressions; the syntax can be anything you like. So, there™s no right

On nature and language

answer to the question: what are the true rules of formation for well-
formed formulas of arithmetic? What are the axioms of arithmetic?
The answer is: any set of axioms you like to generate all the theorems.
It™s the theorems that are real, not the axioms; the axioms are just a
way of describing them, one of many ways. Similarly, if you invent a
computer language, it doesn™t really matter which rules you pick to
characterize its expressions; it™s the expressions that are the language,
not the specific computational system that characterizes them. That™s
not the way natural language works. In natural language there is some-
thing in the head, which is the computational system. The generative
system is something real, as real as the liver; the utterances generated
are like an epiphenomenon. This is the opposite point of view.
Furthermore, the semantics of natural language and of formal
languages seem to be totally different, at least in my opinion. Unlike the
observation about syntax, which is a truism, this thesis is controversial.
Not many people agree with me about this, but in my opinion they are
totally different. In a Fregean formal system, or in any special-purpose
system that anyone would construct, the symbols are intended to pick
out things, real things. That™s an ideal for natural sciences too. If you
construct a scientific theory you want its terms to pick out real things
of the world. I mean, if we postulate Empty Category Principle (ECP),
we™re assuming there™s something in the world which corresponds
to ECP, that is the purpose of the subject. Scientists may also talk
about longitude, let™s say, but they know it™s not a real thing, it™s just a
notation for describing things. But it™s a goal for science “ and it™s built
into every invented symbolic system “ that the terms pick out some-
thing: that™s their semantics, the word“thing relation, essentially.
Now, it™s a real question whether natural language works like that.
I don™t think it does. In that case it deviates even in this respect from
invented symbolic systems. In fact, it seems it deviates at just about

An interview on minimalism

every essential point, and you have to ask why does language have
these properties; it is a fair question. A lot of the questions, I think,
are too hard, like if it™s true, as I believe, that there™s no word“thing
relation, the question why there is no word“thing relation is at the
moment too hard.
But other questions may not be, like morphology. So let™s ask
the question why language has morphology, why should language
have this apparent imperfection? The primary issue concerns one part
of morphology. For example, plurality on nouns is not really an im-
perfection. You want to distinguish singular from plural, the outside
systems want to know about that. So, in fact, plurality on nouns is
rather like different words: just as you have “table” and “chair,” you
have singular and plural, and there are sensible reasons why plural
should be an inflection and “chair” shouldn™t. Namely, everything has
to be singular or plural, but not everything has to be a chair or not
a chair. So there are plausible reasons why some part of morphology
should be there. Formal languages don™t do it but they are just not
interested in singularity and plurality, that™s not an interesting differ-
ence. But human language is interested in this difference, so it has it,
like a lexical item, and languages express it as an inflection because
of its generality in the system, as distinct from “table” versus “chair,”
which is not generalizable. So that part is not an imperfection. What is
an imperfection is plurality on verbs. Why is it there? You already have
it on the noun, so why do you have it on the verb, or on the adjective?
Inflection for number looks redundant there, and that is an imper-
fection. To put it differently, that feature, or that occurrence of the
feature, say, plurality on the verb, is not interpreted. You only interpret
it on the noun, and that™s why in traditional grammars it was always
said that the verbs agree with the nouns and that the adjectives agree
with the nouns, not conversely. Actually, until very recently from the

On nature and language

point of view of generative grammar or structuralist grammar, agree-
ment just looked like a relation. There is no asymmetry to it, no sense in
which verbs agree with nouns any more than nouns agree with verbs,
one would have thought. And as we know, if you look superficially
at languages, it may look as if it is the agreement on the verb that
counts, as in Italian, a Null Subject Language. It looks like it™sthe inflec-
tional features of the verb that are conveying the information, not of the
noun. In fact, there are functionalist studies that reach that conclusion.
If you submit these questions to the minimalist critique, things
look quite different. It looks as if there is some real truth to the tradi-
tional idea that verbs agree with nouns and not conversely. The thing
that is agreeing, presumably the verb, the adjective, the article, and so
on, they all seem to have uninterpretable features, features that are not
independently interpreted by the outside systems. So, what are they do-
ing there? That™sthe imperfection. The imperfection is uninterpretable
Agreement features are an interesting case, because sometimes
they are interpretable and sometimes they are not. But another interest-
ing case is in fact Case. Case systems and inflectional systems have been
studied for thousands of years. That™s the core of traditional grammar,
inflectional systems including Case systems, there™s a huge literature
on that. By the 1940s and 50s it was getting pretty sophisticated within
the structuralist framework. So, say, Roman Jakobson™s “Kasuslehre”5
is a sophisticated interpretation of Case systems. But as far as I can
determine, there was never any distinction made between what we
now call Structural and Inherent Case; I don™t know the literature well
enough to check, but I asked other people like Giuseppe Longobardi,
and apparently there is no clear recognition of the distinction. In
Jakobson™s “Kasuslehre,” he crucially doesn™t make a distinction;
his intent is to show that every feature has all the “right” properties

An interview on minimalism

(as in the standard structuralist approach), so that each Case feature
must have semantic properties. So, Ablative has a semantic property,
etc. Then he tries to show that also Nominative and Accusative have
real semantic properties. But, well, they don™t. There™s a split between
the Cases that have semantic properties, like, say, Dative, mostly, and
the ones that don™t, like Nominative and Accusative (or Ergative and
Absolutive). As far as I am aware, this split was not noticed until the
P & P approach came along; then it suddenly emerged very quickly, in
the early 1980s, that this core system of natural language, which had
been studied for centuries, in fact millennia, broke up into two parts,
one of which is an imperfection (at least prima facie) and the other
which is not. So, the inherent Cases, the ones which are semantically
associated, are really not an imperfection: they are marking a semantic
relation the interpreter has to know about (like plurality on nouns). On
the other hand, why do we have Nominative and Accusative (or Ergative
and Absolutive), what are they doing? They are not interpreted: nouns
are interpreted exactly the same way whether they are Nominative or
Accusative, and that is like inflectional features on adjectives or verbs:
it looks as though they shouldn™t be there. This does lead to interest-
ing questions. If you are interested in the minimalist questions, what
you™ll ask is exactly that: why are they there? I think there is at least a
plausible suggestion: they are there as perhaps an optimal method of
implementing something else that must be there, namely dislocation.
The semantics of expressions seems to break up into two parts,
at least: what was at one time called Deep and Surface Structure
interpretation. It seems there are just different kinds of semantic
properties: exactly how they subdivide is not entirely clear, but you
can see some differences. There™s the kind that have to do with what
are often called Thematic Relations, such as Patient, Experiencer, etc.;
and there™s the kind that look discourse related, such as new/old

On nature and language

information, specificity, Topic, things like that. They seem to be differ-
ent categories of semantic properties, and how to make the break is not
very clear. Take quantifier scope; in the work of twenty-five years ago
that was taken to be the prototypical surface property, now it is taken to
be the prototypical non-surface property, LF-property. It™s not obvious
from the unanalyzed phenomena. But as you learn more, you do see
things breaking up into different kinds and then, within the architec-
ture of a more articulated theory, they even seem to appear in different
places, assuming the theory is right. So there are the LF-related prop-
erties and there are the more surface-related properties. If you look
at the surface-related properties, they are typically edge phenomena,
they have to do with the edge of the construction. So, say, specificity
is typically indicated at the edge of an expression (take Object Shift
for instance, a kind of movement to the edge of verbal phrases which
yields specificity, old information, etc.). And there is a tradition, which
is hard to make clear, but certainly has something to it, which holds
that the surface subject tends to be more or less specific; there are ex-
ceptions, but it tends to have the specific interpretation. That™sperhaps
the same point. Real Focus is also an edge phenomenon, in the Left
Periphery, and all of these things seem to have in fact some peripheral
character. On the other hand, the other category of semantic proper-
ties seems to be non-dislocated, not at the edge; rather, it involves local
relations to other elements that assign the semantic property; a Noun
Phrase is related to a verb, a preposition or something like that. That
gives the Theta relations. If that™s the way the thought system works,
there are two kinds of information it is looking for: one edge related,
the other locally related. Then, well-designed languages are going to
have a dislocation property. An expression will somehow have to dis-
tinguish these kinds of information and in fact an optimal way of doing
it would just be to resort to dislocation; expressions are phonetically

An interview on minimalism

interpreted at the edge even though they are semantically (themati-
cally) interpreted at the local position, the position of Merge. That™s a
plausible reason, external reason, as to why languages have the dislo-
cation property.
Now, you have to implement the property somehow. How do
you implement it? Several things have to be indicated, to make it
work. Now we are internal to the computational system. It™s as if we
had assigned an engineer the problem, “implement the dislocation
property,” because the system has to do it. So, how do you do it? You
have to find the target of dislocation, and it looks as if everything is
driven by heads, so let™s assume that. If you find a target of dislocation,
which will be some head, you have to identify it by some property,
which will also determine what kind of element it attracts to it: a Noun
Phrase, an interrogative phrase, something else? Furthermore, that
head has to make available a position of dislocation; some do, some
don™t. And you have to find the thing that is dislocated. So, there must
be three things: you need three properties, in technical terms, three
features; the term “features” just means properties that enter into the
computational system. So, the engineer recognizes: “OK, I need three
features”: a feature that will identify the target and determine what
kind of expression can move to it, one that will identify the thing that
is to be dislocated, and one that will decide whether the target has
an extra position or not. In fact, the thing that is moved is identified
by Structural Case, the target is identified by redundant features “
Agreement features if it is attracting a Noun Phrase “ and the extra
position is the EPP feature. What has always been considered weird is
the Extended Projection Principle (EPP), “extended” because there is
no semantic role involved; the role is “here™s a position to which you
can dislocate,” where an element can be interpreted as dislocated. So it
seems that you need three features and you have three uninterpretable

On nature and language

inflectional features; this suggests, at least, that the uninterpretable
features are there precisely to implement dislocation.
There™s more evidence for that. One of the properties of the
computational system is that, minimally, it has to satisfy the inter-
face condition: expressions have to be interpretable at the interface.
You can™t have things at the interface that the other systems cannot
read. For example, at the sensorimotor level you couldn™t have a word
that wasn™t spelled out phonetically because the sensorimotor sys-
tem would not know what to do: you couldn™t have an orthographic
word, for example. And the same is going to be true at the thought
end: you have got to eliminate the uninterpretable features. So, some-
how the computational system is eliminating all these uninterpretable
features, but how will it eliminate them? The natural answer is to elim-
inate them once they have done their job. If their job is to implement
dislocation, then, when they have done it, eliminate them. And it looks
as if that is the way in which things work. So, once these features have
done their job, they can™t do it again: once structural Case has been
satisfied, you can™t satisfy it again somewhere else. With agreement it™s
a little more tricky, because there are internal reasons why the system
seems to be doing it many times, but once you have taken care of an
agreement feature, it can™t agree with something higher, for example.
It is frozen where it is. All these things hang together in such a way as to
lend some plausibility to the idea that these are not imperfections, they
are part of an optimal way of satisfying an external requirement, the
interface conditions. I don™t think this is a knock-down argument. It™s
a plausibility argument but it has some force, and if that is right, then
the inflectional morphology turns out to be not an imperfection. Parts
of it, like plurality on nouns, are extremely natural, it™s good design;
other parts like, say, structural Case or agreement features on other

An interview on minimalism

elements, seem to be doing a job that the computational system must
carry out and it is a good way of doing it, in fact.
Now, that good way of doing it does lead to oddities: so, for
example, sometimes the uninterpretable inflectional morphology
functions even though there is no dislocation, with unaccusatives,
for example. Suppose we find a structure with a target T that has both
(redundant) Agreement features and an EPP feature, but the phrase
that agrees with T is unable to move to the target because something
else satisfied the EPP feature: perhaps an Expletive, as in (1), or a phrase
that is closer to T and therefore preempts the displacement by virtue of
locality conditions, as in (2), where t marks the position from which
the phrase to-me raised to the subject position, satisfying EPP:

(1) There T-seem (to me) to be many people in the room
(2) To-me T-seem t to be many people in the room

In English, the rule forming (2) is blocked, but not in other languages;
for example Icelandic, or in such Italian constructions as A Gianni
piacciono i dolci, in line with your analysis of experiencer verbs.6 In such
cases, we have “long-distance agreement” of T and the nominal phrase
that remains in its initial position, many people in examples (1) and (2)
(or i dolci, in the Italian experiencer construction). Visibly, many people
and i dolci agree with the target T (hence indirectly with the verb that
adjoins to T). But according to the account sketched here, the Case “
Nominative Case “ is also assigned as a reflex of this agreement; in
some languages, such as Icelandic, the presence of this Case is also
visible. In such examples as these, we have all the elements that enter
into displacement, but the agreeing nominal is not dislocated. This is a
result of blind operation of the mechanisms “designed” to implement
displacement, blocked here because other factors intervene.

On nature and language

In case (2) the mechanisms do apply but not to the elements that
manifest agreement; rather to the target T and to-me, the latter with
inherent dative case, expressing a semantic relation that is independent
of the Case-Agreement system. Other considerations, still more theory
internal, suggest that there is also a kind of “Agreement” between
T and the closer raised dative, accounting for the local displacement
to satisfy EPP, but only partial agreement, hence not manifested, in
accord with general principles.
This is the research direction: try to show that the apparent
imperfections in fact have some computational function, some opti-
mal computational function. And there are other cases to be thought
about. One massive case is the phonological system: the whole phono-
logical system looks like a huge imperfection, it has every bad property
you can think of. Consider the way an item is represented in the lex-
icon, with no redundancy, including just what is not predictable by
rule. So the lexical item does not include the phonetic form in every
context, if that is predictable by rule; it just includes what the phonol-
ogy must know in order to give the output, and it™s a very abstract
kind of representation, abstracted from phonetic form. Probably none
of the elements that appear in the lexical representation are inter-
pretable at the interface, that is, they are all uninterpretable features.
The interface is some kind of very narrow phonetic representation,
maybe not even that, maybe a syllabic representation or a prosodic
representation. The prosody is not in the lexical item, therefore it is
added along the way; what is in the lexical item couldn™t be read at
the interface, it has to be modified along the way. Probably the entire
phonology is an imperfection. Furthermore the phonological system
has, in a way, bad computational properties. For example, one reason-
able computational optimality condition is the Inclusiveness Condi-
tion, which holds that the computation shouldn™t add anything new; it

An interview on minimalism

just takes the features that it has and rearranges them; that is the best
system, it doesn™t add junk along the way. The phonology violates it,
wildly. The whole narrow phonetics is new, metrics is new, everything is
just added along the way. If you look at the phonetics, it seems to vio-
late every reasonable computational principle that you can think of. So,
that raises a question: is the phonology just a kind of ugly system? Or
is it like what inflectional morphology might be, that is, the
optimal solution to some problem? Well, there is a problem that the
phonology has to satisfy, that an engineer designing the language
would have to address. There are syntactic structures being generated,
and they are being generated the way they are to satisfy the LF condi-
tions, the thought conditions; there is a sensorimotor system, it has
its own properties. The syntactic structures have to interact with this
“external” system. So, the engineer would be forced to find some way
of relating the given syntactic objects to the given sensorimotor system.
It would be nice to show that phonology is an optimal way of doing
it. That™s a meaningful question, maybe way too hard, but certainly a
meaningful question. The best answer that you could hope for is that it
is an optimal way of doing it. I suppose that some day it will be possible
to turn this into a realistic question, a real research question. A ques-
tion like this doesn™t even arise until you think of it in these terms, but
once it arises it makes a lot of sense, and in fact everything in language
can be looked at in this way. The fact that there are parameters ought
to follow from something; why didn™t the system just have one state
that it could achieve? Why these parameters and not others? There is

<< . .

( 19)

. . >>

Copyright Design by: Sunlight webdesign