Hierarchical Network Model of Semantic
Memory
(Collins &
Quillian, 1969)
Basic
Ideas
1. Information is
stored in categories.
2. Categories are
logically related to each other in a hierarchy:
Broad categories of information, like “animal”, are subdivided into
narrower categories, like “bird” and “fish”, which in turn are subdivided into
still narrower categories.
3. “Cognitive
Economy”: Information stored at one
level of the hierarchy is not repeated at other levels. A fact is stored at the highest level to
which it applies. For example, the fact
that birds breathe is stored in the ANIMAL category, not the BIRD category.
How
Information About Animals Would Be Organized
|
|
|
 |
 |
||
This is also a theory about retrieval from LTM. To study retrieval, subjects are given
simple statements and are asked to respond “true” or “false” as quickly as
possible. Researchers measure the
“reaction time”, the time between presentation of a statement and the response. This is called a “speeded verification task.”
Reaction time has a long history in experimental
psychology. Generally, it has been used
as a measure of the complexity of mental processes. The assumption is that the longer it takes
you to respond to a stimulus, the more mental steps you had to go through to
make that response.
Suppose the statement is, “A canary can sing.” When you hear, “A canary”, this activates the
canary category in memory. You then scan
the properties of the canary category for relevant information. If you find it, you stop the search process
and respond. In this case, you would
respond “true”.
Suppose the statement is “A canary has wings.” You start by performing the same steps as
before, but you don’t find relevant information. So you follow the line up to the next
category, BIRD. You then scan the
contents of the category for relevant information. You find “has wings” in this category so you
stop the search and respond “true”.
This is 2 steps more than you had with the previous
statement. Mental steps take time to
perform. Your reaction time should be
longer than it was to “A canary can sing”.
Suppose the statement is, “A canary eats.” You go through all the steps you did with the
previous statement plus 2 more: move up
one level of the hierarchy to ANIMAL, then scan the properties.
The retrieval process is similar to a form of logical
deduction called a syllogism. In a
syllogism you are given two premises and then a conclusion. The first premise is a general principle,
like “All humans are mortal.” The second
premise is a specific case, like “Socrates is a human.” Then comes a conclusion: “Socrates is mortal.”
With the statement, “A canary eats”, it’s like you think,
“All animals eat. A canary is an
animal. Therefore, a canary eats.”
Do we use deductive processes in retrieval? True or False: ABRAHAM LINCOLN HAD A PANCREAS.
How did you know? Did
you learn that specific fact in grade school when you went over the
presidents? And how did you know that
canaries eat? That’s the kind of
reasoning that the model assumes takes place when you retrieve information from
the hierarchy. You don’t just recall
facts. You put facts together logically.
Experiment
Subjects were given a number of statements requiring recall
from different levels of the hierarchy.
The sentences related to a variety of topics, not just animals, but will
be illustrated with sentences starting with “A canary...” There was a mixture of true and false
sentences. Only results for true
sentences will be presented.
(see graph – 1)
You can also measure the time it takes to scan a
category. Present a statement that only
requires the subject to get to a category but not scan it, like “A canary is a
bird.” Compare that to “A canary can
fly.”
The reaction time for the second statement should be longer
because it requires one extra step:
looking at the properties stored in the BIRD category. This prediction was supported.
(see graph – 2)
One thing that literally doesn’t add up here is that the
reaction time for “is a bird” at Level 1 is shorter than the reaction time for
“can sing” at Level 0. The Level 1
sentence should take longer because it includes Level 0’s steps plus the extra
step of going to Level 1.
There are several major predictions of the theory that
turned out to be wrong. Each of these
represents an effect that a theory of semantic memory must somehow explain.
|
“A canary...”
 |
 |
||
1.5 can fly
 |
can sing
Mean
Reaction
Time
1.2
(Seconds)
1.0
 |
0
0 1 2
Levels of
True Sentences
|
“A
canary...”
|
|
||
1.5 can fly
|
can sing
Mean
Reaction
is an
animal
Time 1.2 is a bird
(Seconds)
1.0 is a canary
|
0
0 1 2
Levels of
True Sentences
Typicality effect: All instances of a concept are not equally
good examples of it.
Familiarity effect: Familiar terms are verified faster than
unfamiliar terms regardless of their position in the hierarchy.
Direct concept-property
associations: This is the most important
because it violates the assumption of cognitive economy. Properties are associated with each category
in the hierarchy, not just the highest category.
Problem
Sample Sentences Model
Predicts Finding
Familiarity A. A bear
is an animal. B faster A faster
Effect B.
A bear is a mammal. than
A. than B.
Typicality C. A
robin is a bird. C = D C faster
Effect D.
An ostrich is a bird. than D.
Concept- E.
An animal breathes. E
faster E = F
Property F.
A bird breathes. than
F.
Associations
Network models
Collins and Quillian (1969) proposed that semantic
knowledge is underpinned by a set of nodes, each representing a specific
feature or concept, which are all connected to one another. Nodes that related
in some way, such as often coincident in time, are more strongly connected.
For example, in the model developed by
Collins and Quillian (1969), each node represents a specific word, such as
"bird". Each node is stored together with a set of properties, such
as "has wings" or "can fly. Furthermore, in this model,
connections link categories to exemplars, representing a hierarchical
arrangement. For example, "bird" is connected to "chicken".
Features, such as "can fly" are stored only at the category level,
such as "bird", in which they represent key properties.
To retrieve this knowledge, some cue or stimulus
activates one set of nodes, which then activate other related nodes, called
spreading activation. To illustrate, in response to the question "Is a
chicken a bird", the time to answer this question depends on the number of
connections that intervene between the node that represents chicken and the
node that represents "bird".
Collins and Loftus (1975) then refined this
model. They weighted the connections to explain the typicality effect-the
finding that typical instantiations of a category are recognized more rapidly.
Nevertheless, this model cannot explain a finding, observed by Glass, Holyoak,
and Kiger (1979), that individuals can readily respond to questions that are
patently false, like "Is a chicken a meteor". In this instance, the
nodes are far apart, but the responses are rapid.
Later, more sophisticated network models were
developed (for an example, see Cravo & Martins, 1993). These models are
similar to the propositions proposed by Collins and Loftus (1975). However,
each node might represent some other element, like a concept or feature, rather
than merely a word. Furthermore, the links or connections can represent a
variety of relationships, not just hierarchy.
Collins and Quillian’s Network Model
An older model still in use today is that
knowledge is represented in terms of a hierarchical semantic (related to meaning
as expressed in language—i.e.,
in linguistic symbols) network. A semantic network is a web of elements of
meaning (nodes) that are connected with each other through links (Collins &
Quillian, 1969). Organized knowledge representation takes the form of a
hierarchical tree diagram. The elements are called nodes; they are typically concepts. The
connections between the nodes are labeled
relationships. They might indicate category membership
(e.g.,an “is
a” relationship
connecting “pig” to “mammal”),
attributes (e.g., connecting “furry” to
“mammal”), or some other semantic relationship.
Thus, a network provides a means for organizing concepts. The exact form of a
semantic network differs from one theory to another, but most networks look
something like the highly simplified network shown in Figure 8.2. The labeled
relationships form links that enable the individual to connect the various
nodes in a meaningful way.
Spreading Activation Model of
Semantic Memory
(Collins &
Loftus, 1975)
Like Collins & Quillian’s
hierarchical network model, this theory says that long-term memory contains
interconnected units of information.
These connections produce associations between the units (you think of one,
you automatically think of the other) or pathways that control how you retrieve
information (you must travel along the connecting lines).
Whereas Collins & Quillian
say that the connections were based on logic (set-subset relationships),
Collins & Loftus say that the connections are based on personal experience
and are not necessarily logical.
Additional features of the spreading activation model are as follows.
1. Concepts and properties are treated equally
in the sense that each can be accessed directly. In Collins and Quillian, properties are
contained within concept categories: To
think of a property, like “can fly”, you first have to think of a category,
like “bird”.
2. Not only are properties linked to concepts,
but also to other properties. For
example, “can fly” could be linked directly to “can sing”. In Collins & Quillian, each of these is
linked to a category.
3. Links between units of information vary in
length. The longer the line between two
units, the weaker is the degree of association between them.
Advantages of the model are that
it can explain the familiarity effect, the typicality effect, and direct
concept-property associations. It’s
biggest advantage is that it explains “priming”: you are more likely to retrieve information
from memory if related information (the “prime”) has been presented a short
time before.
The disadvantage is that you
can’t predict reaction times in a verification task until you have mapped out
the individual’s network of associations. The theory explains a lot but
predicts very little.
Examples
|
|||||
|
|||||
|
|||||
|
|
|||
 |
|||
|
|
|
Suppose you are presented with the following statement in a
verification task:
A bird is an animal.
When you hear “bird” and “animal”, this activates the
elements in memory that correspond to those words. The activation spreads out along all lines
connected to the elements. In order to
respond “true”, the activation from each element must meet on the same
line. For comparison, consider the statement,
An insect is an animal.
The line connecting insect to
animal is longer; it’s a weaker
association. So it will take longer for
the activation to meet on the line, and the reaction time will be longer.
This would be the typicality
effect. For this individual, an insect
is a less typical member of the animal category than is a bird.
Note that the property “breathes”
is connected both to Animal and Bird. In
Collins and Quillian, it would be connected only to Animal.
Note the longer line between
Aardvark and Animal than between Canary and Animal. It should take longer to verify, “An aardvark
is an animal” than “A canary is an animal.”
This would be an example of the familiarity effect.
No comments:
Post a Comment