TEACH SEMNET1 Tom Khabaza
12th October 1985.
This is the first of a pair of teach files describing semantic networks.
Semantic networks are a technique for storing knowledge in a computer; this
teach file tells you something about the theoretical nature and significance
of semantic nets; its companion file, TEACH * SEMNET2, shows you how to use
them in your own POP-11 programs.
To understand this teach file and its companion,, it is helpful if you already
understand the concept of a "database".
You will be familiar with this concept if you have already read ANY of the
following teach files:
TEACH * LONDON, TEACH * DATABASE, TEACH * RIVER2.
Contents of this file:
-- Semantics - brief digression
-- Semantic nets are slightly old-fashioned
-- Why study Semantic Networks?
-- What are Semantic Networks?
-- Semantic nets are association-based
-- Labelled arcs: different kinds of associations
-- Distinguishing the roles in the association
-- Structured kinds of association
-- Each node represents a concept
-- Semantic nets and Logic
-- Use of knowledge in Semantic Nets
-- Spreading activation
-- Problems with spreading activation
-- Inference using semantic nets
-- ISA links
-- ISPART links
-- CONNECTS links
-- The many uses of semantic nets
-- Dangers with semantic nets
-- Semantic networks are not a canonical representation scheme
-- The meaning of a node
-- Type-token confusion
-- Readings
-- Bibliography
-- Semantics - brief digression ---------------------------------------------
Before explaining the idea of a semantic network, I will digress briefly to
explain the word "semantics". The term "semantics" refers to the study of
"meaning", but various more specific studies come under this heading. In
particular, a Linguist might be quite happy to talk about the "meaning", and
therefore the semantics, of a sentence in English as being some piece of
logic, whereas a Philosopher might wish to ask about the "semantics" of logic.
So we can ask about the "meaning" of various different kinds of expression,
for example expressions of English or expressions of Logic.
This difficulty with the word "semantics" is typical of discussions in
Cognitive Science. In a unified discipline most practitioners will agree
broadly on the meaning of the terms used, but in an inter-disciplinary area
such as Cognitive Science, the intended meaning for a technical term (like
"semantics") can depend on which of the related disciplines (e.g. Philosophy
or Linguistics) one is most familiar with.
However, the use of the term "semantic" in semantic networks is relatively easy
to explain. Semantic nets were originally developed (by Quillian, 1968) to
describe the meaning of words. Thus, although they form part of the
repertoire of techniques for knowledge representation in Artificial
Intelligence, they are related to semantics by their origin.
-- Semantic nets are slightly old-fashioned ----------------------------------
The most important work on semantic networks was done in the late '60s and
early '70s; they have not been the focus of a great deal of research in recent
years (with the exception of the development of "partitioned" semantic nets,
not described in this file).
For this reason, most of the original readings on semantic nets are from
before the "knowledge based" era in AI; they tend to be expressed less in
terms of knowledge based systems, and more in terms of conceptual structures
and semantic memory (both terms taken largely from Psychology).
-- Why study Semantic Networks? ---------------------------------------------
There are two particularly good reasons for studying semantic nets. Firstly,
many classic pieces of theoretical AI were based on semantic nets; Woods'
(1975) "What's in a link?" paper is a good example of this.
Another important reason for studying semantic networks in an AI course
is that they provide a simple knowledge representation scheme and inference
mechanism to start with. In their simpler forms, semantic nets provide a
knowledge representation scheme that is easy to understand, and inference on
this kind of knowledge is a relatively simple process.
It is an important fact that, especially with more complex schemes for
knowledge representation, the behaviour invoked by a large knowledge base may
be quite difficult to understand. In this respect, knowledge "engineering"
(that is, writing a knowledge base) is rather like computer programming, in
that the larger and more complex the knowledge base, the less predictable is
its behaviour.
-- What are Semantic Networks? ------------------------------------------------
In the following sections, I will describe the major features of semantic
nets. The successive sections develop an example, representing the fact that
canaries are yellow. Each section shows a slightly more developed version
than the previous one, by adding another feature characteristic of semantic
network representations.
-- Semantic nets are association-based ---------------------------------------
The notion of association of ideas is an old one in Psychology. Semantic nets
capitalise on this; the basic notion in semantic nets is that of an association
between two concepts. Each concept is represented by a "node", shown by an o
in the figure below. Different concepts are distinguished by having different
labels on the nodes; the words "canary" and "yellow" in the figure are the
labels. The association of two concepts is shown by drawing a line (or "arc")
between the relevant nodes; these are shown as a line of hyphens in the
figure.
Figure 1:
canary o-------o yellow
This network represents the fact that there is some connection between
canaries and yellow.
This net has five components: two NODES (shown as o), two LABELS, the words
"canary" and "yellow", and an ARC, shown as a line of hyphens.
-- Labelled arcs: different kinds of associations ---------------------------
In addition to simply associating two ideas, we may want to distinguish
between different KINDS of association. This is done by labelling, not only
the nodes, but also the arcs; thus different kinds of association are shown by
giving the arcs different labels.
Figure 2:
colour
canary o----------o yellow
This network represents the fact that the relation between canaries and yellow
is something to do with COLOUR.
-- Distinguishing the roles in the association -------------------------------
In addition to showing the KIND of association we want to represent, we may
want to make the associated concepts play different ROLES in the relation.
For example, we may want to express the fact that "the colour of canaries is
yellow", rather than that "the colour of yellow is canary".
To do this, we can give an arc a DIRECTION; an arc with a direction is called
a DIRECTED ARC. In the figure below, the direction of the arc is shown by an
arrow symbol ">".
Figure 3:
colour
canary o---->-----o yellow
This network represents the fact that the colour of canaries is yellow.
-- Structured kinds of association ------------------------------------------
Finally, in addition to giving an arc a simple word as a label, we may wish to
show that the concept used to label the arc also has some relations to other
concepts, and is thus also a node. For example, we might wish to show that
there are different ways in which a thing may have a colour; it may have that
colour only on the surface, such as the yellowness of a canary, or it may have
that colour all through, such as the greenness of grass.
Figure 3 shows this distinction, by making the label on the arc between
"canary" and "yellow" be "surface colour", which has the "is a kind of"
relation to the abstract idea of "colour".
Figure 4:
colour o--------<-------+
| |
^ ^
is a kind of | | isa
| |
o |
surface colour |
canary o---------->---------------o yellow
This network represents the fact that the surface colour of canaries is
yellow, that surface colour is a kind of colour, and that yellow is a colour.
-- Each node represents a concept -------------------------------------------
When we draw a semantic net, the intention is that the nodes represent
CONCEPTS. (Note that "concept" is being used here as a Psychological term;
semantic nets have been used by Psychologists as a model of semantic memory,
for example, Collins & Quillian (1969), Anderson & Bower (1973), Anderson
(1976)).
So in the above examples, nodes represent CONCEPTS, arcs represent RELATION
between concepts, and the label on an arc shows you WHAT relation it
represents. Finally, the fact that the label on an arc can also be a node
simply means that a particular relation is also a concept.
-- Semantic nets and Logic ---------------------------------------------------
Semantic nets are often compared to logic, for two reasons:
Firstly, because historically, both were intended to relate to the "meaning"
of language. (Remember that, as mentioned above, semantic nets were invented
to describe the meanings of words.)
Secondly, semantic nets are easier to understand than logic. This is clearly
true of the simple forms described above; the pictorial representation of
semantic nets can seem to capture a concept in a very intuitive way. This
becames less true with more complex cases, as described below. In these more
complex cases, semantic nets have to be extended in a way which makes them
much less easy to understand.
-- Use of knowledge in Semantic Nets -----------------------------------------
In this teach file, I will describe two kinds of use of knowledge represented
in semantic nets. The first is very general, and is used more to explore the
structure of some existing facts than to deduce new facts. The second is much
more specific, and can be used, given certain kinds of facts represented as a
semantic net, to deduce facts not previously given. In many ways, both
techniques are only knowledge use of the most elementary kind. However, they
do illustrate the general principle that knowledge represented as a semantic
net can be USED by a program to some specific end.
It is the second, more specific kind of technique, that we will be using for
programming in TEACH * SEMNET2.
-- Spreading activation ------------------------------------------------------
The "spreading activation" technique was first introduced with semantic nets
by Quillian (1968). The description below is not exactly like Quillian's
program; whereas below, I will describe the processing of a whole sentence to
find its "conceptual structure", Quillian's program worked only on pairs of
words to find their conceptual similarities or links. The description below
is informal, and takes ideas not only from Quillian's work, but also later
work such as Anderson & Bower's (1973) "Human Associative Memory".
The idea of the spreading activation technique is that, when we hear a
sentence, the nodes in our "internal semantic net" corresponding to the
important words in the sentence are "activated". This simply means that those
nodes are in some way "marked" as being relevant to the sentence. The process
then proceeds as follows: every node that is linked in any way to any active
node is also activated, in a way that will show, if we examine the node, which
of the words in the sentence caused it to be activated. This process is
repeated until some node is activated by MORE THAN ONE of the original words;
this indicates that a link between them has been found. We may continue the
process until we have found all the links that we can reasonably expect to
find, or until we can be sure that we have found all the links shorter than a
certain length. At whichever point we choose to stop, we should expect to
have at the end of the process a collection of nodes and the links between
them, which represent the "conceptual structure" of the sentence.
Let's look at an example:
Suppose that we start with the sentence "The ship sank beneath the sea". Look
at the network in figure 5, below; the initial sentence would activate the
nodes "ship", "sink", and "sea". (We will not worry here about the difference
between "sink" and "sank"; also, to simplify the example, we will leave out
the activation of the "beneath" concept, which labels one of the arcs.)
Figure 5:
float
o
/ | \
/ | \
can > ^ > on
/ | \
/ | \
ship o opposite o water
\ | / |
\ | beneath |
can > V > ^ isa
\ | / |
\ | / | contains
o o----->------o fish
sink sea
Now, each of these "active" nodes will activate its neighbours. The "ship"
node will activate the "float" and "sink" nodes; note that here we have
already got a node (sink) that has been activated twice; once because it was
in the sentence, and once because it was linked to "ship" which was in the
sentence. The "sink" node will activate the "ship" node (again) and also the
"water" node, which will also be activated by the "float" node. The "water"
node will also be activated again by the "sea" node.
By this time, each of the "float", "sink", "ship", "water" and "sea" nodes has
been activated several times. There are also some other nodes (e.g. the
"fish" node in figure 5) activated, but not as many times. The "conceptual
structure" of the sentence is then considered to be the collection of nodes
that have been most often activated, and the arcs between them.
-- Problems with spreading activation --------------------------------------
Experimenting with this technique can produce interesting results; however, it
also has a number of problems. Among them is the fact that identical
conceptual structures can be formed from completely different sentences;
the above example would have worked exactly the same way if the sentence had
been "The sea sank beneath the ship". From the point of view studying natural
language, this problem is caused by taking no account of syntax when analysing
a sentence.
In general, the problems of the spreading activation approach are caused by
the fact that it is in many ways "unprincipled" - it is an attempt to use
semantic nets in an ad-hoc way, without worrying too much about what it all
means. This has the result that the technique is UNRELIABLE; that is that
although it can produce correct results, it can also produce incorrect ones.
The techniques described in the next section are of a different kind.
Although more limited in their application, the are far more reliable. Barring
various specific pitfalls, some of which will be described later, we can
assume that if the original net contains only correct information, then using
these techniques will only ever produce correct results.
-- Inference using semantic nets --------------------------------------------
In Artificial Intelligence, when we talk about "inference", we are referring
to reasoning with some given knowledge. That is, a person or a program, given
some knowledge, and the ability to use their knowledge in certain ways, can
come up with new knowledge that in some sense FOLLOWS from the knowledge
originally given.
In the domain of semantic nets, there are various kinds of reliable inference.
The following sections describe three of them, which are summarised in the
table below. These are also the kinds which are used in the programming
examples of TEACH * SEMNET2. If you are unfamiliar with the technical terms
(like "transitive" and "commutative") used in the table, don't worry; these
will be explained in the relevant sections of this file.
Table 1: 3 reliable types of inference with semantic nets.
Type of link Characteristics
------------ ---------------
isa "isa" hierarchies, property inheritance.
ispart transitive relations.
connects transitive relations, commutative relations.
The next three sections explain and illustrate each of these in turn.
-- ISA links -----------------------------------------------------------------
These are arcs in a semantic net marked with the label "isa", for example:
Figure 6:
isa
bird o--->---o animal
This represents the fact that a bird is a kind of animal.
Links of this kind, that is, links representing that one concept is an
instance of another concept, can be used to reason in the following way:
If an A is a kind of B
and a B has property P
then an A also has property P.
Here is an example of this kind of reasoning:
If a bird is an animal
and every animal has a heart
then a bird also has a heart.
When we make use of our knowledge in this way, we say that we have "made
an inference". The above inference would be made using a net like this:
Figure 7:
isa
bird o--->---o animal
|
V has
|
o heart
Here is a more complicated net involving more "isa" links, with more possible
inferences:
Figure 8:
isa bird isa animal isa
canary o------>------o------>------o------>------o organism
| | |
isa | | |
pigeon o------>------+ V has V has
| |
| |
o o
wings heart
>From this net (figure 8) we could infer the following facts:
Canaries have wings.
(because a canary is a bird, and birds have wings)
Pigeons have wings.
(because a pigeon is a bird, and birds have wings)
Birds have hearts.
(because a bird is an animal, and animals have hearts)
And also:
Canaries have hearts.
(because a canary is a bird, and birds have hearts)
Pigeons have hearts.
(because a pigeon is a bird, and birds have hearts)
These last two facts, that canaries and pigeons have hearts, were inferred by a
slightly more complex method form of the same method. We had to use not only
facts that we were originally given, but also a fact that we had INFERRED,
namely that birds have hearts.
When we have a semantic net which has many "isa" links, so that many objects
are a kind of some other object, we say we are using an ISA HIERARCHY. Figure
8 is an example of an isa hierarchy.
We would call things like "having wings", or "having a heart", PROPERTIES of
birds, canaries etc. The process of inferring that something has a property
because it is a kind of something else which has that property is called
PROPERTY INHERITANCE. We say that the concept "canary" INHERITS the property
of having wings from the concept "bird".
Property inheritance is an important concept in AI. I will not discuss its
other uses here, but simply note that it is used widely, both in the design of
knowledge based systems, and in more mundane AI programming.
-- ISPART links --------------------------------------------------------------
These are links marked with the "ispart" label, as in:
Figure 9:
ispart
wheel o---->-----o car
This net represents that fact that a wheel is (or can be) part of a car.
Links of this kind, that is links that represent a fact about one kind of
object being part of another kind, can be used in the following kind of
reasoning:
If A is part of B,
and B is part of C,
then A is part of C.
For example:
If a hub is a part of a wheel,
and a wheel is a part of a car,
then a hub is also part of a car.
This inference might be made using the following net:
Figure 10:
ispart ispart
hub o---->-----o wheel o---->-----o car
Here is a slightly more complicated net:
Figure 11:
ispart wheel ispart
hub o------->-------o------->-------o car
|
ispart |
rim o------->-------+
Using this net we can infer the following facts:
A hub is part of a car.
(because a hub is part of a wheel, and a wheel is part of a car)
A rim is part of a car.
(because a rim is part of a wheel, and a wheel is part of a car)
This kind of relation, where if A has some relation to B, and B has some
relation to C, then A has the same relation to C, is called a TRANSITIVE
relation. The "ispart" relation is not the only transitive relation; there
are many other common ones, a good example being "isin":
If the block is in the box,
and the box is in the house,
then the block is in the house.
However, the main relation of this particular (transitive) kind we will use in
this teach file and in TEACH * SEMNET2 is the "ispart" relation.
-- CONNECTS links ------------------------------------------------------------
These are links marked with the "connects" label, as in:
Figure 12:
connects
Victoria o--------------o Green Park
This net represents the fact that there is a connection between Victoria (a
station on the London underground), and Green Park (another station on the
London Underground). In this example, and all of those of TEACH * SEMNET2, I
am talking in the domain of the London Underground system; so if I mention a
station, I mean an underground station. I will also talk about "connection",
which in the example will mean underground train connections.
Links of this kind, that is links which indicate a CONNECTION between two
objects, allow two kinds of reasoning (or inference). Firstly, they are
TRANSITIVE, like "ispart" links. Thus
If A is connected to B
and B is connected to C
then A is connected to C.
Secondly, they are COMMUTATIVE, that is:
If A is connected to B
then B is connected to A.
A commutative relation is one which "goes both ways". Many relations are not
commutative, for example "ispart" - if a wheel is a part of a car, this does
NOT mean that a car is a part of a wheel. But with "connects", if A is
connected to B, then B is connected to A.
Here is a copy of figure 12:
connects
Victoria o--------------o Green Park
Note that there is no arrow on the "connects" arc. The arc is UNDIRECTED,
i.e. it has no specific direction. Commutative relations like "connects" are
shown either by undirected arcs, or by arcs showing an arrow in BOTH
directions, as in:
Figure 13:
connects
Victoria o--<-------->--o Green Park
indicating that the "connects" relation "goes both ways".
Here is a slightly more complex example with connects relations, using
stations in the London Underground map:
Figure 14:
connects connects
Victoria o------------o Green Park o------------o Bond Street
In this net we would make inferences to the effect that everything in it
connects to everything else, that is:
Victoria is connected to Green Park.
Green Park is connected to Victoria.
(because Victoria is connected to Green Park)
Green Park is connected to Bond Street.
Bond Street is connected to Green Park.
(because Green Park is connected to Bond Street)
Victoria is connected to Bond Street
(because Victoria is connected to Green Park and
Green Park is connected to Bond Street)
Bond Street is connected to Victoria
(because Victoria is connected to Bond Street
and also because Bond Street is connected to Green Park and
Green Park is connected to Victoria).
So the "connects" relation has two important properties: it is TRANSITIVE,
that is it carries over from one object to another, and it is COMMUTATIVE,
that is it "goes both ways".
-- The many uses of semantic nets -------------------------------------------
As mentioned above, semantic nets have been of interest to psychologists as a
model of "semantic memory", see for example Collins & Quillian (1969),
Anderson & Bower (1973), Anderson (1976); in fact they were invented by
Quillian as a model of human semantic processing.
However, semantic nets have been used for a variety of knowledge
representation tasks. A famous example of such a use is Winston's learning
program (see, for example, the last part of Winston's (1977) chapter 2).
Winston's program was given example descriptions of a type of object build out
of blocks, and had to learn the distinctive characteristics of the object,
represented in the form of a semantic network. The example always used is the
concept of an ARCH.
Figures 15 and 16 show examples of arches; the kind of characteristic that
Winston's program had to learn was that the "lintel", that is the object at
the top of the arch, could be any object (e.g. a pyramid); it did not always
have to be a rectangular block.
Figure 15: An arch with a block for the lintel.
##############
##############
##############
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
Figure 16: An arch with a pyramid for the lintel.
##
########
################
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
@@@@ %%%%
Figure 17 shows the kind of semantic net that the program had to build as a
description of the concept "arch".
Figure 17: Semantic net description of an arch.
o object1
/ | \
/ | \
on < V > on
/ | isa \
/ | \
object2 o o object o object3
\ | /
\ | isa /
isa > ^ < isa
\ | /
\ | /
o block
Figure 17 describes an arch as made of three objects, one of which is any
object, and resing on two others which are blocks. Winston's actual networks
were more complex that this; his book shows a number of examples.
-- Dangers with semantic nets ----------------------------------------------
Semantic networks have been important in the history of AI; however their use
presents problems which to some extent account for their present unfashionable
status. The following sections describe some of these problems.
-- Semantic networks are not a canonical representation scheme -------------
Some AI workers had hoped that semantic networks would provide a "canonical"
representation of meaning (for example, for natural language sentences). The
idea of a CANONICAL representation is that, for example, if we had two
different sentences that meant the same thing, they would have the same
semantic network representation.
For example, if semantic networks were a canonical representation scheme for
the meaning of English then the sentences:
John loves Mary
and
Mary is loved by John
(assuming that you believe that these sentences mean the same thing), would
have EXACTLY THE SAME semantic network representation.
However, it turns out that this is not the case; semantic network
representation schemes are sufficiently vague that many different networks can
be found for any given sentence, let alone for two apparently different
sentences (see Woods, 1975 for a discussion of this issue). For example, the
sentence "John loves Mary" might be represented as either of the networks in
figure 18:
Figure 18: 2 semantic nets for "John loves Mary".
loves
John o----->-----o Mary
o Loves
/ \
/ \
source V V target
/ \
/ \
John o o Mary
It should be noted here that the same problem is encountered with all
knowledge representation schemes; it is arguable that a canonical scheme for
knowledge of any useful sort is impossible. The problem with semantic
networks is not simply that they are not canonical, for neither are the
systems that replaced them. Rather, the problem is that semantic networks
were being used as though they WERE canonical; once it was realised that they
were not, they did not seem so attractive.
The fact that knowledge representation schemes are seldom (if ever) canonical
has various implications in AI; on the positive side, it means that when using
a given scheme, there is always hope for finding a better representation of a
concept than that which you are using. On the negative side, it means that no
matter what representation you use, your system may always have the same
concept in different forms, and never be able to detect the fact. Arguably,
this is also true of human beings!
-- The meaning of a node -----------------------------------------------------
A semantic networks is a STRUCTURAL description of a concept. That is, at
least theoretically, the important part of the net is its structure rather
than, say, the labels on the nodes and arcs. However, if it is the structure
rather than the nodes used that is important, then labels are in some sense
ARBITRARY; it should be possible to change them without making a significant
difference to the meaning of the net.
Now clearly, if we change, say the "John" label in figure 18 to be "Jack", the
network means something different, however we could argue that since the node
still represents a person, and all we have changed is the name, the concept is
essentially the same.
But this is not always the case; with some some networks, if we change the
labels on the nodes, we change the nature of the concept. Figure 19 shows an
example.
Figure 19: Changing the meaning by changing the label:
isa colour
bird o------<------o------>------ yellow
Tweetie
isa colour
bird o------<------o------>------ yellow
canary
isa colour
bird o------<------o------>------ yellow
yellow-bird
Depending on what label we put on the middle node in this network, the network
represents a SPECIFIC yellow bird (called "Tweetie"), a KIND of yellow bird
(that is a canary), or the CONCEPT of a yellow bird itself (any yellow bird).
Thus it is always unclear what a specific node is supposed to mean. The next
section shows an example of one important type of confusion that can occur
because of this.
-- Type-token confusion ------------------------------------------------------
Suppose we want to represent the meaning of the sentence "The dog is wet".
We might (naively) choose to do so as in figure 20:
Figure 20: The dog is wet.
is
dog o--->---o wet
Now suppose we want to add the additional fact that "Fido is a dog". We might
do so as in figure 21:
Figure 21: Fido is a dog (and the dog is wet).
isa is
Fido o--->---o dog o--->---o wet
But now, due to property inheritance, we would also believe that Fido is wet,
which is wrong.
What has happened? The answer is that we have confused the notion of "dog" as
a TYPE of object with the notion of some SPECIFIC (or TOKEN) dog (the one
which is wet). Thus, the property of wetness, which should have been attached
to the specific (token) dog was instead attached to the TYPE dog, representing
the notion that "All dogs are wet" (and hence, if Fido is a dog, Fido must be
wet too).
This problem is called "type-token confusion", and is something to watch out
for when dealing with (or reading about) semantic networks. The possibility
of this confusion (or at least the importance of making the distinction) was
discovered with semantic nets by their originator (Quillian, 1968), but is
point a point of general importance in all knowledge representation.
-- Readings ----------------------------------------------------------------
This section contains some recommended readings on semantic networks, and
brief notes on them. Full references to the readings can be found in the
bibliography section at the end of the file.
Readings on the basic concepts of semantic networks:
Norman (1978) (O.U. course material).
This contains a very short section on semantic nets.
Barr & Feigenbaum (1982) Vol. 1, Chapter III: "Representation of Knowledge",
Section C3 "Representation Schemes: Semantic networks".
Barr and Feigenbaum's book is an advanced AI textbook. However, if you
can ignore his slightly terse style, this section should be reasonably
readable.
More advanced readings:
Winston (1977)
Chapter 2, last section on "Learning simple descriptions".
A good description of Winston's learning program, as described above.
Rich (1983)
Chapter 7, sections 7.1 - 7.2.1.
This is a good introduction, except that Rich assumes a familiarity with
basic formal logic. Unless you are familiar with it, ignore the pieces of
logic. The main advantage of this reading is that it describes
partitioned semantic nets.
Woods (1975) "What's in a Link?"
Classic paper on the pitfalls of semantic nets, and on the foundations of
knowledge representation in general.
Slack (1978) (O.U. course material).
This is rather long, but worth skimming to get the gist of how semantic
nets are used as models of human memory.
Collins & Quillian (1969)
A report of a classic experiment, looking at the psychological
plausibility of semantic networks. Unfortunately, the results were open
to more than one interpretation.
-- Bibliography ------------------------------------------------------------
After each reference in this section, the University of Sussex Library
classification mark is given in square brackets.
ANDERSON, J. R. (1976) "Language, Memory and Thought", LEA.
[QZ 1030 And].
ANDERSON, J. R. and BOWER, G. (1973) "Human associative memory", Hemisphere.
[BF 371 And].
BARR, A. and FEIGENBAUM, E. A. (1982) "The Handbook of Artificial
Intelligence", Volumes 1 and 2, Pitman.
[QZ 1240 Han].
BOBROW, D. G. and COLLINS, A. (1975) "Representation and Understanding:
Studies in Cognitive Science", Academic Press.
[QZ 1010 Rep].
COLLINS, A. M. and QUILLIAN, M. R. (1969) "Retrieval time from semantic
memory", Journal of Verbal Learning and Verbal Behaviour, Vol. 8, pp
240-247, reprinted in "Offprints booklet", Open University, course D303:
"Cognitive Psychology", 1979.
[BF 311 Ope].
FINDLER, N. V. (1979) "Associative Networks: Representation and Use of
Knowledge by Computers", Academic Press.
[QZ 1220 Ass].
HENDRIX, G. G. (1977) "Expanding the Utility of Semantic Networks through
Partitioning", in IJCAI 4.
MINSKY, M. (ed.) (1968) "Semantic Information Processing", MIT Press.
[QE 100 Min].
NORMAN, D. (1978) "Overview", Open University, course D303: "Cognitive
Psychology", Units 31-32.
[BF 311 Ope].
QUILLIAN, R. (1968) "Semantic Memory", in Minsky (ed.) (1968).
RICH, E. (1983) "Artificial Intelligence", McGraw-Hill.
[QZ 1240 Ric].
SLACK, J. (1978) "Semantic memory", Open University, course D303: "Cognitive
Psychology", Block 3: "Memory (part 2)", Units 18-19.
[BF 311 Ope].
WINSTON, P. H. (1977) "Artificial Intelligence", Addison-Wesley.
[QZ 1240 Win].
WOODS, W. A. (1975) "What's in a link: Foundations for Semantic Networks", in
Bobrow & Collins (1975), pp 35-82.
--- $poplocal/local/teach/semnet1
--- Copyright University of Sussex 1989. All rights reserved. ----------