NETWORK MODELS (TLC - Quillian/ Spreading Activation
- Collins & Loftus)
Semantic memory is a netlike organization of concepts
in memory with many interconnections. Each concept is represented by a NODE
or a location in the network and there are LINKS or associations which
connect NODES. There are SUPERORDINATE LINKS and MODIFIER LINKS
Superordinate links show that the concept is a member of a larger class. LTM
Structure is based on semantic relatedness rather than hierarchies. Semantic
relatedness is reflected in the strength of links connecting concepts.
Modifier links
show the properties of the concept
SPREADING ACTIVATION - naming activates a node,
the activation expands or spreads to other related nodes. Activation spreads
first to all of the nodes linked to the original node and then spreads to
more remote nodes. When a node has been activated it is tagged along with
information about the nature of the activation. Activation will occur
from 2 sources and we get an INTERSECTION (Collins & Loftus, 1975).
If the activation is strong enough, it will be attended to. Frequently
used links have greater strength and therefore activation travels faster between
2 nodes (typicality). The greater the distance between items, the weaker the
spread. The greater the number of irrelevant paths (and interfering information),
the weaker the spread.
Semantic
Distance Effect - RT verification increases with the distance between
S and P terms ("a canary is a bird" vs. "An Ostrich is an animal") but how
about "A ostrich is an animal" vs. "An Ostrich is a bird"? The typicality
effect or relatedness explains the faster RT to the first sentence
In some cases, relatedness does not have the usual
effect of inhibiting RT for false responses (All fruits are vegetables vs.
All fruits are flowers).
Evidence: Meyer & Schvanaveldt (1971)
effects of context on sentence verification/ processing one word impacts
(facilitates or inhibits) the processing of another word - positive and
negative priming (nurse - doctor). Findings: meaning affected speed but
physical similarity (nurse - purse) did not
Neely
(1977) when the prime was the name of the category and the letter
string an instance of the category (automatic spreading activation)
Advantages
of this model:
1) accounts for semantic relatedness effects;
2) shows that the meaning of a word is greatly
affected by associations between the word and its associated concepts;
3) assumes that there is more than one way to facilitate
verification
Problem:
Ratcliff & McKoon (1981) found evidence supporting
the notion that the amount of activation arriving at one Node is a decreasing
function of the number of links that the activation has traversed but did
not find evidence that activation takes a significant amount of time to spread
between two nodes. Activation was present at the same moment at all activated
nodes and so the time required for activation was not a function of distance
from the initially activated node (parallel, simultaneous activation)
SEMANTIC FEATURE MODEL SMITH, SHOBEN, & RIPS
A concept is defined as a set of features (like
properties or modifiers in the network models). Birds have wings, fly, lay
eggs; these are all features of bird. The semantic features of a concept combine
to provide its meaning.
A two stage decision process is used to
make judgments about concepts:
1) comparing all features of the subject of
a sentence and the predicate (A canary is a bird)
may show low similarity > quick false
may show high similarity > quick true
may show intermed. sim. > stage 2
2) compare defining features of both the subjects
and the predicate
2 types of features: defining features -
necessary to the meaning of the item (robin has a red breast)
...................................characteristic
features - descriptive but not essential
Rips
(1973) asked Ss to indicate how closely related each of several
instances was to its category (chicken, sparrow, robin, penguin) and
to other instances of the category. They then translated the ratings into
distances in semantic space (multidimensional scaling).
Accounts for fuzzy set, relatedness, semantic distance
Loftus
(1973) importance of noun order;
Organization
of facts in propositional memory and their retrieval times:
1. if a fact about a concept is frequently encountered,
it will be stored with that concept even if it could be inferred from a
more distant concept;
2. the more frequently a fact about a concept is
encountered, the more strongly that fact will be associated with
that concept and the quicker it will be verified (Birds fly vs. birds walk);
3.
verifying facts that are not directly stored with a concept but must be
inferred takes a relatively long time.
Procedural
memory - knowledge of how to do things, makes use of production
systems (e.g. IF A is the mother of B and B is the mother of C THEN A
is the grandmother of C - a production rule)
A production rule consists of a set of conditions
and an action. The conditions are preceded by IF and the action is preceded
by THEN
Working memory
- contains the information that the system can currently access (information
from LT declarative memory and temporary structures deposited by encoding
and the action of productions). It is declarative knowledge, permanent or
temporary, that's in an active state. This is where the action takes place.
Five
assumptions of ACT*:
1. strength - each link in the network has
a strength; the strength of a new link is low but is strengthened each time
the link is used
2. activation - at any instance a small
portion of the nodes in LTM are in an active state - this is working memory;
the rate ofspread is a function of the strength of that
path relative to the sum of the strengths of all paths coming from that
same concept. The more information associated with a concept, the longer
it takes to retrieve a particular association (e.g. If you know 3 things
about A and only one thing about B, then it will take longer to retrieve
a piece of information about A) - FAN EFFECT
3. spread of activation - activation spreads
out from the active nodes to passive nodes to which it is linked
4. dampening - periodically activation is
dampened throughout the network; occurs as a function of time
5.
active list - a maximum of 10 nodes can be kept on the active list
(working memory)
Connectionist
Models (Parallel Distributed Processing)
(McClelland, Rumelhart, & Hinton, 1986)
An attempt to develop a model of higher level processing
based on our understanding of neural processing. PDP models assume "
that information processing takes place through the interaction of a large
number of simple processing elements called units, each sending excitatory
and inhibitory signals to other units"
Marr - neural computation - How might the brain achieve higher processing? Start
from general knowledge of how neurons work and ask - howcould higher level functioning be achieved by
connecting together basic elements like neurons?
PDP set up a network representing knowledge where
each element or instance unit is a neuron linked or connected together
(excitatory connections). Retrieval is the activation of the unit corresponding
to the item which activates all the properties for the item thus creating
a pattern of activation and the greatest activation leads to retrieval.
Inferences are based on similarity
Cognitive processes involve interactions among
these units (excitation or inhibition). Many interactions occur simultaneously
or in parallel. Knowledge is not stored in specific nodes. Knowledge
consists of connections among simple units distributed throughout the network
. Knowledge is encoded in connection strengths.
Connectionist schemes can represent information
without recourse to symbolic entities like propositions; they represent
information sub-symbolically in distributed representations, they have the
potential to model complex behaviors. They hold out the possibility of theories
of cognition that map directly onto the neurological substrateThe concept Rose in a distributed model does
not have symbols that represent the rose specifically but rather stores
the connection strengths between units that will allow either the sight
or scent of the rose to be re-created (accounting for imaging and propositional
experiences).
The sight and scent of the rose can be viewed as
being coded in terms of simple signals in certain input cells (vision units
and olfaction units). The network is capable of associating the pattern
of activation which arrives at the vision units with that arriving at the
olfaction units. The distributed representation of the sight of the rose
is thus represented by the matrix of activation over the units in the network,
without recourse to any explicit symbol for representing the rose.
Distributed representations are content-addressable
meaning that any part of a past occurrence or scene can lead to retrieval
of the whole memory. Distributed representations allow automatic generalization
meaning that patterns that are similar will produce similar responses. past
occurrence or scene can lead to retrieval of the whole memory. Distributed
representations allow automatic generalization meaning that patterns that
are similar will produce similar responses.