2. One of the most basic distinctions in the psychology of
language that concerns both the representation and the
acquisition of language competence is the symbolic versus
connectionist contrast. Hulstijn (2002a) explains that symbolic
accounts represent knowledge as a collection of symbols
accompanied by rules that specify the relationship between
them. By symbol we mean a unit that exists independently of
the context. We can also think of a radically different type of
architecture of knowledge- connectionist account- whereby
knowledge is not represented as sums of tiny informationpacked units but rather as activation patterns in a neural
network.
3. The distinction between nativist and non-nativist
approaches refers to whether the infant is assumed to
bring a genetically coded linguistic endowment to
language acquisition or not. The standard claim is to
equate Nativism with symbolism and non-nativism
with connectionism.
Chomsky’s prominent symbolist theory of generative
grammar is explicitly nativist, whereas for
connectionists the idea of innate predeterminism with
no process explanation seems to be at odds with
current theories of the development and the
functioning of the brain.
4. We are born with a set of rules about
language in our brains.
“Children are equipped with an innate
template or blueprint for language and
this blueprint aids the child in the task of
constructing a grammar for their
language.” (Chomsky 1965)
This is known as “Innateness Hypothesis.”
5.
6. Now the question is, “What is UG?”
It's a theory in linguistics credited to Chomsky, proposing
that the ability to learn grammar is hard-wired into the
brain. The theory suggests that linguistic ability manifests
itself without being taught and that there are properties
that all natural human languages share. It is a matter of
observation and experimentation to determine precisely
what abilities are innate and what properties are shared by
all languages. It contains the core principles underlying any
human language and the parameters for any allowable
variation that these languages can manifest. UG is not a
theory of language acquisition because it does not spell out
how the biologically driven process of developing a nativelike system of grammar takes place.
7.
8. The poverty of stimulus paradox (also known as the
"logical problem of language acquisition" or "Plato's
problem"), concerns the fact that learners come to
know more about language than they observe from
experience. A child may acquire a language even
though the data itself is too poor to determine the
language: the child needs no evidence for much of the
knowledge she brings to the learning situation. Since
the child can fixate on any language in the face of a
poverty of stimulus about each language and since all
languages are equally acquirable, children all begin
with the same universal linguistic knowledge.
This is the essence of the poverty of stimulus
argument.
9. Another frequently support for UG is the ' no negative
evidence in input argument'. Much of the debate around
the Poverty of the Stimulus Argument focuses on
negative evidence. If there is lot of negative evidence
there are more chances that the child’s learning is based
on trial and error. Fortunately while making
mistakes, children are neither corrected nor do they pay
attention to the corrections of adults.
It also makes one wonder about the existence of a
predetermined language faculty when considering how
universal and uniform the L1 acquisition process is.
Crain and Thornton (2006) list several more subtle
points in support of UG. Most of these are highly
technical and follow the pattern that a specific linguistic
structure or rule is so complex that it could not possibly
have emerged spontaneously as a product of
experience, without some innate regulations.
10.
Lack of neurobiological validity
The innatism issue
Speed of language acquisition
The selective nature of UG
Grammaticalization and other experience-
based processes.
11. UG admittedly does not
aim to account for language
acquisition
processes, because as
Lightbown and Spada
(2006) declared, several
proponents of UG openly
admit that the theory is not
a good explanation for the
acquisition of a second
language, especially by
learners who have passed
the critical period or who
receive formal L2
instruction.
Even if one rejects the
strong form of nativism that
presupposes some built-in
language-specific processing
module such as UG, there
might be a case for a weaker
form of form of nativism
that would state that some
initial human disposition
for language learning is
required.
12. This view has been gathering outside the
mainstream of linguistics in the fields of cognitive
science, psycholinguistics, and developmental
psychology and has been increasingly applied to
language acquisition and recently to SLA. Within
this broad movement there are five interrelated
but distinct theoretical approaches: (1)
Connectionism; (2) The competition model; (3)
Dynamic system theory; (4) Emergentism; and (5)
Usage- based theories.
13. Connectionism refers to a broad
approach in cognitive science that
utilizes computational modeling as its
main research tool and aims at
explaining various mental processes
within the framework of simple
interconnected units (nodes).
14. As N. Ellis (2006a) summarizes, 'Connectionist simulations are
data-rich and process-light.' The essence of connectionist
modeling is to devise systems that are able to find patterns in
rich linguistic stimuli fed into the model with only a few
relatively uncomplicated algorithms. A successful connectionist
model can exhibit rule-like behavior even though no rules are
explicitly represented and no symbols are passed around in the
network.
One of the most powerful determiners in a connectionist
architecture simulating language acquisition is the frequency of
the various elements in the language input. Items that are
frequent in the input increase the connection weights between
the nodes in the network, and the output of such models can
show that seemingly rule-like behavior can emerge simply from
the processing of input, as a function of the consistency and
frequency of various input properties.
15.
16. A connectionist system (often called a neural
network) consists of a large number of processing
units (i.e. nodes) that are linked together in some way.
There are at least three types of nodes: input
nodes, which receive information, output nodes, which
represent the outcome of the model's processing
work, and hidden nodes, which are within the
network, between the input and output nodes, often in
various layers. Depending on the nature of the
nodes, neural networks can vary greatly.
17. In the purest form of
parallel distributed
processing (PDP), the nodes
themselves do not represent
any distinct piece of
information because it is the
pattern of activation that
carries knowledge. In this
network type, any piece of
information is made up of
the contribution of many
processing units (hidden
nodes) contributing to it.
Another connectionist
principle is the localist
network. A connectionist
network in which
single, dedicated units
represent distinct pieces of
information without any
hidden nodes in between.
Such a system can be
connectionist if it allows for
the spreading of activation
from one unit to the next in
a parallel manner.
18. The incremental adjustment of the connection weights
in a network to reach a certain level of input accuracy is
called error correcting or supervised learning. The most
common form of this learning type is back
propagation(which is an abbreviation of backward
propagation of errors.) and the bulk of the work in the
modelling of language processes has used this approach
(MacWhinney 2001c). Backpropagation involves a stepwise
process, whereby the output of the network is compared to
the desired output, and the error for each output neuron is
calculated.
A comprehensive neural network simulating language
processing might include hybrid- that is, both symbolic
and subsymbolic- elements, resulting in a combination of a
parallel distributed and a localist network.
20. The core of the competition model is a
distributed connectionist network that links the
form of an input sentence to the function of that
form, that is, its meaning in context. In
deciphering the sentence function, the listener/
reader is unconsciously looking for various cues in
the input stream.
In this model there are multiple cues that
compete with each other, the winning clue
determines the perceived meaning of the sentence.
21. In their experiments MacWhinney and Bates (1994)
specified the following four cue types: (1) the linear
surface position of the argument vis-à-vis its head; (2)
the affixes attached to the argument; (3) the affixes
attached to the head; and (4) the inherent lexical
semantic features of the argument. A key
characteristic of the competition model is the
assumption that linguistic knowledge is 'probabilistic
all the way down'. Cue strength depends entirely on
the cue's reliability and availability in the input.
Thus, the competition model offers a theoretical
explanation of how the distributional properties of the
input control language learning and language
processing.
22. With this model, MacWhinney (2005,2008) made an
attempt to broaden the scope of the competition
model, which was not designed to give a full account of
SLA, so that it can cover the acquisition and processing
of both L1 and L2 in a unified model. There are two
major differences between competition and unified
model: (1) several new components were added to the
paradigm; and (2) instead of a PDP
network, MacWhinney used a different type of neural
network, a self-organizing map or 'Kohonen map', as
the underlying organizational framework.
23. The construct of cue cost is added to the model, referring
to limited cue evaluation when quick online decisions are
called for.
A further new component is the concept of
resonance, which MacWhinney (2008) considers the most
important area of theoretical development. It was inspired
by the neurobiological understanding of effective
vocabulary learning. MacWhinney argues that resonance
can be used to consolidate new forms on the
phonological, lexical, and construction levels. It is also
thought to be involved in code-switching: if a language is
repeatedly accessed, it will be a highly resonant
state, delaying or interfering with the activation of another
language.
24. the unified model views long-term linguistic
knowledge as organized into a series of self-organizing
maps (Kohonen maps). In these neural
networks, nodes with the highest activation are moved
in the direction of the input pattern, thus creating a
spatial structure that reflects the activation pattern.
Thus these maps generate a representation of the
input sample with additional topographical
visualization.