Knowledge Representation

4 IssuesS
Syllabus
a l Agents: Knowledge-based agents, The Wumpus world, Logic,
onsitional theorem proving. Efective propositional model checking,Propositional
Agents basedlogic,
on
Pro
propositional logic.

et Order Logic: Representation Revisited, Syntax and Semantics of First Order logic, Using First
Orderlogic

Contents
Representation and Mappings... .. Winter- 14, 18, 19,
Summer- 16,18,20
.

Marks 7

4.2 Approaches to Knowledge Representation... Summer 15, 17, 18,20

. Winter-18 .Marks7

4.3 University Questions with Answers

 4-2 Knowledge Rep
Artificial Intelligence sentation IsSU8Sue 4-3
GTU : Winter-14,19, Summer,
A r t i f i c i e lI n t e l l i g e n c e
Knowledge Representation Issues
and Mappings
4.1 Representation
16,18,20 This can be treated as John Zorn plays in the band Naked City or John Zorm's band

4.1.1 Introduction is Naked City.

Another representatic
ation is band = Naked City
Some knowled
programs require
.Search-based problem solving to be John Zorm Bill Frissell, Fred Frith, Joey Barron
or path toward so
band-members
states
can be a particular
implemented. Knowledge
this knowledge must be represented in a na
etc. Before being used ular way Granulartty

Representation (KR) is an important ould

the knowledge be represented and what are the primitives.
with a certain format. Knowledge issue in what
level shoul
in particular, "The dominant naw At
larity of Representation Primitives are fundamental concepts such as
science in general and in Al
-

computer
since the early 1970s has been based on #h
aradigm
for Choosi ng
the
Granularit

and as English is a very rich language with million

over halfa
building intelligent systems seing, playing
knowlec premise
wledge is represente holding, find difficulty in deciding which words to choose as our
that intelligence presupposes knowledge". Generaly, we will clear upon
words
it is
the svstem's knowledge base, which consists of data structures and prograns. a series
of situations.
in
primitives
to have a program called an infa
addition, the intelligent system is expected erence feeds a dog then it could become :

for the task at hand. Th E Tom

engine that implements the reasoning patterns necessary Thus
current Al theory and practice dictate that intelligent systems be knowledge based feeds(tom, dog)
ased, bone like:
gives the dog
a
consistent with this simple knowledge base plus inference engine architecture. This Tom
f
emphasis on knowledge has led to suggestions that AI can be arguably called these the same ?
dog,.bone) Are
gives(tom,
applied epistenology". object food constitute feeding?
In any sense does giving an

> feed(x) then we are making progress.

4.1.2 Issues In Knowledge Representation If give(x, food)
certain inferential rules.
But need to add
we
Are any attributes of objects so basic that they occur in almost every problem How do we represent
Louise is Bills cousin
on relationships
famous program it is Chris
domain ? If there are such attributes then we need to make sure that they are In the
(brother or sister (father mother( bill)) Suppose well. or
handled appropriately in each of the mechanism we propose. If such attributes this? louise daughter =

female and then son applies

as
as a male or
do not know if it is Chris and
exists, what are they ? There are several issues that must be considered when then w e
different levels of primitives
levels of understanding require
representing various kinds of real-wold knowledge. Clearly the separate similar primitives.
to link together apparently must
these need many rules and the underlying question
Important Atributes storage problem
is a potential
Obviously there
Are there any attributes that occur in many different types of problem? comprehension is needed.
be what level of to handle.
is another issue
There are two instance and isa and each is that is granularity
important because each supports T h e finest level of
knowledge be represented.
For this, o n e
property inheritance. knowledge that needs to to be
level of the knowledge
Granularity is understanding of
There are two important attributes that are of general significance such as ISA and primitive (basic)
should have complete
instances. These atributes are important because they support property in an Isa
represented in the system. Existence
handled are Inverses,
Relationship among attributes must be considered carefully which is need to be attributes. Major
inheritance.
depicting more knowledge. O t h e r significant
issues those
about values,
Single-vahued
knowledge is
for reasoning whose
hierarchy, Technique The set of objects
identified.
Relationshlps to be
artributes are required identified.
stored should
be clearly
-What about the relationship between the attributes of an object, such as, nverses, required to be
existence, techniques for reasoning an and Mappings
consider an example of an inverse in
about values and single valued attributes.
The Techniques of Representation large Nevertheless

13 within.
encountered is
knowledge
that
bandJohn Zom,Naked City) solve a complex problems manipulating
AI used to means
of
DE well as some

as
TECHNICAL O Knowledge thrust for knowledge
PUBLICATIONS An up thrust for knowledge TECHNICAL
PUBLICATIONS
An up
 A r t i h c i a li n t e l l g e n c e 4 5
Artificial Intelligence 4-4 Knowledge Representetion lssue
sues Knowledge Representation Issues
nsider example of the dtilated
required so as to create solutions for new problems. In the representation there
here are two board from which h two
squares, in
Checkerbo Problem. onsider a normal
different entities that must be considered checker

all the remaining squares comers, have been opposite

removed. The
.Facts: truths in some relevant world. These are things that we want to reDrees. is to
cover

exactiy with
donimoes, each of which covers
task
No overlapping either of dominoes on
Representation of facts in some chosen formalism. These are things that
squares

of the multilated board are

top of each other or of
alowed. Can this task be done dominoes
two

the boundary
actually be manipulated.
be done in two levels
over

7"
Structuring of these entities can

The knowledgelevel at which facts are described.

No. black square
The symbol level at whichrepresentation of some objects at the knowledge-level 30
defined in terms of symbols that can be manipulated by are
programs. No. white squara
32
Reasoning program
internal
acts
representation
English Fig. 4.1.2 A multilated checker board
English generation A example follows:
understanding .Checkerboard total contains 32 white squares and 30 black squares.
English .When every domino cover two neighboring squares, a biack one and a whute one,
representation
then first thirty dominos cover 30 black squares and 30 white squares, and leaving
Fig. 4.1.1 Mappings between facts and two white square and zero black domino.
representation
Our main goal is to focus on
facts, representation as well as the .These two black squares can not be adjusted and can not cover remaining domino.
that must exist between the two as
shown in the
two-way mappings
4.1.1 Fig. above. The links in the .It is impossible to cover all 62 squares with 31 one domincs.
are called representation mappings. In representation figure
mappings, there are
Forward representation which
-

maps from facts to representation.

Backward representation which
maps the other way.
One
representation of facts concerns with natural
sentences is that, language (particularly English)
regardless of the representation for facts that we use
may also need to be concerned with an in a program, we
order to facilitate English representation of those facts that in
getting information into and out
mapping functions from English sentences to the of the system. We must also have
going to use and from it back to sentences as
representation which we are actually
shown in the Fig. 4.1.1. For
can use mathematical logic as the example we

sentences below. representation formalism. Consider the

Engus
Tommy is dog. This fact can also be
a

represented in logic as follows: Dog(Tommy)

Suppose also we have logical a

explained below. Using the deductiverepresentation of the fact: all dogs have talls a
Partial covernng
Cover fieds on the board
new
representation object. Usingg an mechanisms of the logic, we may generaecould he
then generate the appropriate backward mapping Flg. 4.1.3 Observation

has a tail Or we can functo

English sentence 1 number of black
is that the
representation of new fact to cause us ommy
to take
make use or u An obser Which can be made the computation
in
covering.
h e s a m e is
true
Squaresc dominoes in the partial
representation of additional facts. some
appropriate aáction or derive
POna to the number of

TECHNICAL PUBLICATIONS An up thrust or knowledge

An up thrust for PUBLICATIONS
knowledge TECHNICAL
Artlficial Intelligence
4-6
entation Issues Artificial intelngence

4-7

which enforces
the number of black squarres to coincide Knowledge Represental Issues
ror the number of white fields, 2) Inherltable knowledge
the inter play befween covered sa
with the white squares, when
investigated uares on the edge is
Relational knowledge made up of
object
in the partial covering associativity like
board and dominoes contained values attribute.
co-relation associated
Representation
to Knowledge All data shoule be organised into
42 Approaches GTU: Summer-15 17, 18, 20, Winter
a
hierarchy of classes.
18 Inherit values from being all members of class.
of knowledge in a particularparticular domai
domain shoulad
A good system for the representation Class must be arranged in a generalization.
possess the following Pproperties individual frame can represent the
the ability to represent all of #h Cuery
collection
Representational adequacy It is kinds of associated with a individual node.
of attribute and its
value
that domain.
knowledge that are needed in
I t is the to manipulate the representational
Inferential adequacy ability ct
Player
in such a way as to derive new structures corresponding to new knowl.
inferred from old.
ledge Sa

Iníerrential efficiency It is the ability to incorporate into the

-

struchs knowledge Cricket

additional information, that can be used to focus the attention of the inferene
mechanisms in the most promising direction. Isa Isa
Acquisitional efficiency Acquiring new information easily.

Two types of approaches to knowledge representation: L batsman L.H.batsman

1) Simple relational knowledge instance instance

2) Inheritable knowledge bablee moni

1) Simple relational knowledge Equal Equa

handed handed
This is the simplest way of storing fact which uses relational method, when every BPL Indore
and each fact about a set of objects is set out
sequentially and automatically in
column. Fig. 4.2.1 Inheritable knowledge
This type of representation is small procedure for inference. Example
I t is used to define
inference engines.
For example Properties of Inheritance hierarchy
Player 1) to be point from object and its value.
Weight Age Play cricket
4) Boxed: to be object and value of attribute any object
ww.roo
Monu 70 Right H.
30 3) It may be also be called slot-and filter structure
Sonu 65 Right H.
Bablee
Algorlthm retrieve
50 29 Left H. 0 retrieve a value for attribute of an instance object.
Soni
45 Right H.
2
1Find object in the knowledgebase
Moni 4 I t h e r e is a value for the attribute, report t a t vea
42 Left H
wwwwwe
w***

Player_info (Monu', 70, 30, right H)

PUBLICATIONS
An up for knowledge
thrust

TECHNICA
TECHNICAL PUBLICATIONS. n tn thou
 4-8
Knowledge kepresentation A r t f i c i e lI n t e l l i g e n c e 4- 9
Artificial Intelligence ISsues Knowled Representation
tail, otherwise go
not then
arams rely
Many programs rely on more than one technique
Otherwise look value of instance,
if
found, report it. Otherwise.
node and srstem - They are used
used in
attribute, if one
is
Database system representing Simple
find a value for the
ISA, found for
the attribute. there which declarative facts and can be said as a set of
Relation Knowledge
is no value search using relations of the
Aatabase systems. Fig. 4.2.2 shows an same sort
example of such systems.
4.2.1 Inferential Knowledge
Player Height Weight
is very useful form of inference, represent the kno..
Bats-Thrown
property
When inheritance
as a formal logic.
ledge Ram 6-0 180
Right-Right
tail (n) Shyam 5-10 170
All cat have tails tx: dog (x)> has a
Right-Right
Veer 6-2 215 Left-Left
Set of rules 6-3
Tarun 205
1) Define require fact.
www.ww..a niivnwe nwnmnt
Left-Right
Semantic nets Semantic nets are useful for
2) Additional statement is varified, true or false. representing inheritable knowledge.
Tnheritable knowledge is the most useful for property inheritance, in which elements
3) Logic provides a powerful structure in relationships.
of specific classes inherits attributes and values from more general classes in which

4.2.2 Procedural Knowledge they are included. Frames also do play a big role in representing this knowledge.
In order to support property inheritance, objects must be organized into classes
Procedural different-different way in program.
knowledge can
explain and classes must be arranged in a generalization hierarchy. Fig. 4.2.2 below shows
Procedural knowiedge clearly differs from propositional knowledge. some additional baseball knowledge inserted into a structure that is so arranged.

.Procedural knowledge basically involves knowing how to do Example:

something.
Procedural knowledge follow implicit learming. Handed
Person Right
4221 Advatages of Procedural
Knowledge
1) Property specific knowledge can be
specified. isa

2) Extended logical inference is

possible.
4.2.2.2 Disadvantages of Procedural Knowledge
Adult
male eign178
isa
1) Consistency: all eigh195
deduction are not always correct. bats Baseball
2) Completeness all cases are not Equal to
to
There
easy represent. handed player
are
multiple techniques for knowledge representation. Different representaion Isa, isa bating 252
formalisms are, average

Rules batting averagPitcher batting average262

106 Fielder
Logic instance
instance

Natural language Pee-Wee-

eam Brooklyn
Chicago team Three-Finger dodgers
Database systems Brown Reese
cubs
Semantic nets hierarchy
4.2.2 Inheritance
Frames Fig.

TECHNICAL PUBLICATIONS PUBLICATIONS -An up thust for knowledge

An up thrust
for knowledge TECHNICAL
 4-10 Knowledge Representation lssues 4-11
Artificial Intelligence
Artifcielntellgence Knowledge Repres
esentation Issues
Lines repr
used in semantic nets/ frames.
ent atribtes Adult maleg
hierarchy is normally
sa atributes of objects. Correct da
and values of uction from
Boxed nodes represent objects Brown is 195 cm. An incor
ig.
could be
4.2.2 could be:

:
height of Three-Finger
height of Three-Finger
178
Brown is cm. 1he structure

semantic network
shown in the
or
deduction
Fig 422
19a

Musician
be also called
collection
a

1S a slot-and-filler
structure. It may Isa
frames.

Predicate logic Jazz Avant Grade

inferential knowledge. Jazz
Predicate logic is used to represent
structure in which to describe relationshine
Logic provides powerful s among instance instance

values.
It can be combined with some other powerful description language with Miles Davis John Zom
hierarchy. bands bands

Production rules
Production rules are useful in representing procedural knowledge. Miles Davis Group Naked City
Miles Davis Quintet Massada
Procedural knowledge is form of operational knowledge which specifies what to
do when. 4.2.3 Property inheritance hierarchy
Flg.
such LISP.
Previously it was done using programming language as
fail.
Otherwise look for a value of instance if none
However it was hard to reasoning with this method hence in AA progra 3.
then report it.
find a value for the attribute and
is represented using production rules. 4. Otherwise go to that node and
procedural knowledge for the attribute.
isa until a value is found
5. Otherwise search through using
Inheritable knowiedge is an important issue
in computer science in
Relational knowledge is made up of objects consisting of Knowledge Representation (KR) for building intelligent
"The dominant paradigm
general and in AI in particular. on the premise
that inteligence
Attributes 1970s has been based
systems since the early in the system's
Corresponding associated values. knowledge is represented
presupposes knowledge. Generally, In addition, the
structures and programs.
We extend the base more by allowing inference mechanisms consists of data
knowledge base, which called an inference engine
that

is expected to have
a program
current Al
Property inheritance ntegent system
necessary
for the task at hand.
Thus
Elements inherit values from being members of a class. piements the reasoning patterns based, consistent
systems be knowledge
Data must be organised into a hierarchy of classes (Fig. 4.2.3). and practice dictate that intelligent This emphasis
EOry architecture.
base plus inference engine
i s simple knowledge
Boxed nodes called applied
objects and values of attributes of objects. has led to suggestions
that AI can be arguably
Values can be objects with attributes and so on. Knowledge
Arrows - point from object to its value.
epistemology be termed the
symbol-manipulation approacn.
neural
eapproach described above may of work in
which another approach,
h i s structure is known slot and filler structure, semantic out non-symbolic
as a newo ically, however, AI grew distributed processing
or

collection of frames. networks (or connectionism or parallel w a s outplayed

by the
but this approach an
The algorithm to retrieve a value for an attribute of instance object representations) played the major role, networks again got
an
80s when neural
until the
1. Find the object in the annpulation approach
knowledge base. imn
important role.
2. If there is a value for the attribute
report it.
thrust for knowledge
- An up
ECHNICAL PUBLICATIONS
TECHNICAL PUBLIGATIONS
Artificial lntelligence Knowled 4-13
4-12
epresentation Is ues ArtificialIntelligence

Knowledge epresentation lssues

is mentioned by Davis (2001p.8138): statistical anals armical neural network node attempts emulate this beha
Afinal approach lysis of large t
of input lines
which are analagous to vior. Each node l
corpora of data. set input symapses in
'activation function' (also known as biological neuron.
a a
has an an
The approaches to KR have parallels in theories of psychology as node also
Each
ho node when fire, similar to a
to 'transfer function), a

epistemology. We will start networks, then o

by considering neural as in which
tells
biological neuron. In it's
symbolic" approaches and finally consider large corpora of data (which h e nction can just be to generate a T if
this activation func
the summed
simplest
form,
is most some value or a '0' otherwise.
Activation functions, input is
related to library and information science, which is concermed concerned wit
with large greater
than
to be
this simple -

in fact to create networks that can do not

useful work, they
however, do
bibliographical and full-text databases). have
have to be more complex, for at least some of
a lImost
m
always the nodes in the
Neural networks
network.

neural networks exist, for in the human

.While biological example, brain, Arti6 feedforward
neural network, which i one ot the more common neural
network
Neural Networks (ANN), are mathematical or computational modela .A
mes, is composed of a set or these nodes and connections. These nodes are
information processing. There is no precise agreed definition amongest research
what a neural network is, but the original inspiration for the
chers ranged in layers. The connections are typically formed by connecting each of the
as to
Was technique nodes in a given layer to all of the neurons in the next layer. In this way every
from examination of bioelectrical networks in the brain formed by neurons is connected to every other node in the next
and node in a given layer layer.
their synapses. In a neural network model, simple nodes (or "neurons
are
. Tvoically there layers to feedforward network an input layer,
are at least three a
connected together to form a network of nodes, hence the term "neural network it is
an output layer. The input layer does no processing
-

a hidden layer and

Input Hidden Output vector is fed into the network. The input layer then feeds
simply where the data
ayer layer layer
in turn, feeds into the output layer. The
into the hidden layer. The hidden layer,
network occurs in the nodes of the hidden layer and the
actual processing in the
Input:#1 output layer.
network can be
When enough neur are connected together in layers, the
Feedforward networks, in
Input #2 trained' to do useful things using a training algorithm.
to do inteligent
particular, are very useful, when trained appropriately, Neural
Output tasks on unfamiliar data." (Wikipedia :

classification or identification type

Input#3 network, 2005).
is a neural network much like
As an approach to knowledge representation from
American psychology
Input # 4 Behaviorism dominated
behaviourism in psychology. behaviour of animals
to shape the
interest was how kinds ot
about 1913 to 1970. Its main with different
such organisms
and human beings by confronting approach (or
Fig. 4.2.4 A model of a neural net thus very
much an input-output and to
It is
Sumuli-patterns. mental terms (eg memory)
tried to avoid
(copied from Government web site: - r e s p o n s e approach).
They between stimuli
and responses (eg
ane to relations behaviourists neglected
http /smig.usgs.gov/SMIG/features_0902/tualatin_ann.fig3.gif usGS-authored
or produced them with terms referring most
place response"). Although
a "black
box,
with "delayed the brain as
and information are in the public domain)
to
place "memory" preferred to look at of neural
processes and the idea
I n a typical neural network, each node operates on a principle simila LStructures and brain models, and Hebb
has a in Donald 0.
biological neuron. In a biological neuron, each incoming synapse of a neuro some beha iourists were interested the
behaviourist

forward for the

first tim by
associated with it. When the of each synapse, times
its
inpu networks was put
weight weight
Summed up for all incoming synapses and that sum is greater
in 1949 behaviourism are closely
nets and British
n neural classicai
threshold value, then the neuron fires, sending a value to anotner Doth the computer technology of in particular by
ideas developed
network. epistemological

PUBLICATIONS-
An up
thrust for knowledge
TECHNICAL

TECHNICAL pIURIICATIONo 1 I thnict for knowledge

 Knowledge Representation Issues A t f i c i a lI n t e l l i g e n c e
4 - 15
Knowledge Representation Issues
4-14
Artificial Intelligence
be that knowledge
is
is
represented
represented in the
brain as on the
other hand, reverses this dependency by identifying the meaning of a

The basic
idea may
based orn of
repetitions of similar
sim stimulia knowled base with its Kramer and Mylopoulos, 1992, p. 746).
empiricism. learning
stimuli-processes,
why or view the most imDor
result of point
association.
From our
co issye b) Semantlc networks

follows the laws of is provided by somebody

who is in
D)
representation
what represents wanted
behas emantic networks are knowledge representation schemes involving nodes and
is that knowledge her view ot The
learning process.
It is his or
considered true,
relevant and important
knowled
that inks between nodes. nodes
between nodes. The
represent objects or concepts and the links
links directed and labeled. Semantic
represent relations
what is are
indirectly manage formulated and provided
directly, but is implemage
by cognitive motivated models of human memory.
nets were originaly
is not n
This knowledge
manipulating
the stimulation of the
tem
sres
the system or the organism
by or
Kramer and Mylopoulos (1992, p. 747-748) their popularity and
which involves rewards and/or punishm ent). ccording to
organism (simplified: by feedback uccesscan best be understood as a comvenient compromise between the

Symbol representation declarative and the procedural extremes, while "others have argued that semantic
to knowledge representation
in AL which
networks offer a fundamentally different representational paradigm that object
can : is
There are several approaches be

seen as subcategories of the symbol-representation

They all share the
approach.
centered in the sense that it is based on object descriptions rather than arbitrary
the use of some kind of
conditions the knowledge is explicated by symbolic
propositions and focuses on krnowledge organization.
installed in the system "manually, piece by piece. The four mos
language and is
WordNet is an example of a semantic net. The semantic web is a concept that
based systems b) semani
kinds of symbolic systems may be a) logic tic
important represents a major
research program with semantic networks. For many persons
networks and c) frame-based systems.
this idea of a semantic web represents the kind of knowledge organization with
a) Logic based representations the most promising prospects.
.Knowledge may be represented in computers by programmers writing declarative
sentencessuch as "Socrates is human" and if somebody is human, then she is

dOe-up-of a-player-in
Market
mortal" using mathematical logic. "A major advantage of many logics adopted for
knowledge representation is that they are sound and complete, which means that
derivability and provability lead to the same set of consequences, given a carries
Retailer
knowledge base. It has however turned out to be difficult to find logics that is em
both expressively adequate for knowledge representation and also computationally

ractable. "Atempts to find an acceptable compromise to the expressiveness versus

tractabilitytrade-off generally use variations of first-order
logic, following one of
a-form-of Brand
Creates

The first approach limits the expressiveness of the language of Various

wo approaches.
representation by restricting the form of the formulas that can be admitted in
other Manufacturer

the attributes
s
knowledge base. The second approach redefines the provability relation or na
first-order logic to make it
computational tractable. S Size
Relational databases.. . widely used to
represent "simple" facts, such
peop as
is-a Category

cresses or salaries, constitute Color

(Kramer and
a
good example of the first approach *
Mylopoulos, 1992, p. 745-746). The second
make a slight change in the semantics of existential approach whichexa
may for
Segment
quantification
arge representations computational tractable, but this has on is-a
the provability relation. a remarkable
inpa na

Logic based systems may also use Category

atribute
representations treat the intended procedural representations. ion
that
imposes constraints on meaning of a knowledge base as a
Flg. 4.2.5 Semantic Nets
knowledge base operations. Procedural ntations,
TECHNICAL repe An up thrust for knowiedge
PUBLICATIONS An up thrust for TECHNICAL PUBLICATIONS
knowledge
 4 6 present ntation Issues ArtificialIntelligence9
4-17
Artificial Intelligence wledge Representation Issues
aspects of KR in computer
Frame-based representations that . se frames, a eplstemic
science
are knowledge
representation systems
genera
General
The
R [Knowledge Representation)
architectures we have considered above
Frame-based systems (1975), their primary means to
their
as
primary
notion originally
introduced by Minsky representnt ic-based systems,
[logic-
semantic networks and
frame-based systems],
A frame is
structure for representing a concept
other proposals of a more or less similar together with
uatio
a
domain knowledge. restaurant. Attached to a frame many flavour, such as production
what may be called the classical, or
systems, constitute
in a
are several
such "restaurant" or "being (with some
heaging) the knowledge-based approach to AL. Knowledge representation,question
as

information, for
kindstoofuse
instance,
definitional and descriptive informatio and in this
how the frame. Frames are supposed to capture the essence af.concep
view, involves large, complex structures of symbols, defined and assembled by
or
for dinner, by clustering all
for example going out hand. This approach to Al essentially derives from a line of philosophical thought
stereotypical situations, This means, in particular, that
vant
information for these situations together. running from Descartes through Leibniz, Frege, and Russel. In the late 1980s and

deal of procedurally expressed

knowledge should be part of the a t 1990s, however, as a result
of the inherent difficulty of this line of research, and of
be organized in frame systems in whick
frames. the limited progress that has been made, this approach to AI has been challenged
Collections of such frames are to ich the
frames are interconnected. by two alternative methodologies neural networks, and statistical analysis of
(Davis, 2001, p. 8137-8138).
are in many ways similar to object-oriente large corpora".
Obviously, frame-based systems
of knowledge representation thus correspond to cognitivism in
programming languages; indeed, the two theories interacted strongly in eir The symbolic forms
and to rationalism in epistemology while neural networks correspond
development. psychology both
in psychology and to empiricism in epistemology. They may
The chief advantages of frame-based architectures are expressivity, flexibility and to behaviourism
the subjective side of knowledge representation.
In the symbolic
ease of use. The chief disadvantages is lack of precision and lack of a well defined be said to ignore and
of the programming tasks are defining
form of KR the person in control
model of inference. The architecture provides a wealth of features and options for is said about whether different subjects
both representation and inference, but only a weak underlying model. Hence, in a assembling the knowledge. Nothing would or should define and
different traditions or paradigms)
complex case, it is difsicult to predict how these features will interact or to explain (representing networks the person in control
of
assemble different kinds
of knowledge. In neural learm.
should
unexpected interactions, which makes debugging and updating difficult. what the organism or the system
the stimulation is determining connected to subjective
views
criteria may be
From a psychological point of view has a tendency to overuse frames as is said about how persons' kind of
Nothing assumed without any
socio-cultural factors.
In both c a s e s it is of
explanations been critized: "I going to argue against the existence
am not and to
representation is "objective".
What kinds

(whatever that may be) of organised knowledge structures. What I will do is place examination, that the knowledge quotation:
"More
uncovered by the following
may be
doubts on the explanatory value of concepts as frames, conventions, scripts and so perspective that are missing level has relativistic properties.
that the knowledge observer's
there are structures like frames and scripts, recently, Clancey (1991) argues n v i r o n m e n t . It is
an
on...
Even if are they relatively easy description is of an agent in its
e

studied" (Clancey,
1992,
for people to override. People can still use arbitrary knowledge of the world to A knowledge-level the agent being
possessed by level (or
representations
system's knowledge
understand sentences and scenes: you cannot exclude
any part of the knowledge theory, not
description of an agent's or a
has relative properties
and implies
base in advance, using some general prestructuring of that knowledge. 743). Yes! The
Thereroe generalized: a description of its knowledge)
perspectives.
This is a basic point
the content of such structures as frames and
scripts must themnselves be Dos t or system
from specific
but it has yet
to be fully
analyzable and subject to reasoning by their users, which puts us back a i o the theory of the agent
understanding
of knowledge
where we started. What we have gained is a summary of the agere in pragmatic
the representation. much
theories of knowledge there has not been
implemented in
regularities frequently or typically, exhibited. The structures themselves te epistemology theories and
recognition
of AI as applied hand
epistemological
nothing about people's cognitive capacities, only about what are probably n spite of the between on
the one
This is odd,
because

ephemeral habits of thought which people can change. In terms of Bilg (1987) systematically
investigation
of knowledge
representation.
theory
of knowledge
theories and any
frames and other hand
scrips lack any kind of 'witcraft. Frames, scripts and relatea Onthe
theory ofknowledge
summarize some of the patterms that emerge when people don't bother to ** is the knowledge.
PIstemology a theory of
be based
on
must
(Vliet, 1992) epresentation

knowtedge
thrust for
A n up
PUBLICATIONS
TECHNICAL
 4- 19
Knowledge Representation Issues
A r t i c i e lI n t e l l i g e n c e

4-18
owledge Representatio Issues
Artificia! Intelligence nresent. In large corpora of texts
voice present.
(such as Davis, 2001
and Kramer many voices are present (what kind
representation
in Al
has been es varies accordin8 to how the text corpus is of
n overviews of knowledge and
rationalism
cted as
voice

or scholarly papers).
selected, e.g. if it consists of
Mylopoulos,
1992) only
empiricism
to expand
such overviews b newspapers
and need
There is an
obvious Large corpora of texts onsist of documents each of
which is itself a
approaches.
different
epistemological
positions.
arguments and
knowledge claims. We are now in the realm of Library system of
cOverage of feld of A, first and fore
oremost by and
Hermeneutics has been regarded in the
contributions nclude Mallery, Hu
include
ormation Science (LIS) rather than in computer science in a narrow
sense. What

Winograd &Flores
(1987).
Additional
and Martin (2005). There seems
Tepresented
are
in LIS are
representations of documents
representing
knowledge
and Fonseca
Chalmers (1999) (thus meta-representations). If, tor example, the text
corpus is an academic corpus
Duffy (1992);
bee a need for a
more direct application
ot
eutical/pragmati
historicist/hermeneutical/prao

from the same domain as the person

doing the representation (e.g.
computer
representation. science) then different suggestions and voices on how best to perform the task at
approaches to knowledge
such an approach to knowlad hand is present in the very material to be (meta) represented. Different paradigms
shall very briefly suggest
.

In the next section we of the person dni

main point will
be put on the subjectivity in KR contain arguments in favour of specific ways to do the representation.
representation. The introduced in the first part of #h
the representation in line
with the thoughts this
In other words: The texts to be organized are voices, which probably will contain
article. different implications for how this knowledge should be organized (and by the
Analysis of large corpora way also implications for how texts should be selected in the first hand). This
both neural nets and symbolic KRs E. Davis brief argument may be expanded also to cases in which the corpus is not in the domain
Among alternatives to
or explicit criteria of
mentions text corpora of knowledge representation: Any document has implicit
relevance, which are of importance for organizing those documents.
. "The statistical approach to AI involves taking very large corpora of data and
for how best to represent arts
them in great depth using statistical techniques. These statistics can then If we consider the domain of Arts then the criteria
analyzing As discussed by Ørom (2003)
be used to guide new tasks. The resulting data, as compared to the is depending on what is considered (good) art.
for how arts should be
extremely shallow in terms of their semantic different traditions in Arts have different implications
knowledge-based approach, are
not just the programmers voice.
content, since the categories extracted must be easily derived from the data, but represented. In corpora there are different voices,
and provide knowledge representations
The programmer may ignore these voices
they can be immensely detailed and precise in terms of statistical relations. consider those voices and
exist that based on his own voice alone, or the programmer may
Moreover, techniques such as maximum entropy analysis allow a which represents a dialog between
himself
collection of statistical indicators, each individually quite weak, to be combined provide a knowledge representation, text based on
it is possible to use corpora
effectively into strong collective evidence. From the point of view of knowledge and the voices in the corpora. This way
epistemologies (see also Hjorland
representation, the most interesting data corpora are online libraries of text pragmatic epistemologies rather than empiricist
Libraries of pure text exist online containing billions of words; libraries of and Nissen Pedersen, in press).
extensively annotated texts exist containing hundreds of thousands to millions or
words, depending on the type of annotation. Now, in 2001, statistical methods or Answer in Brief
natural language analysis are, in general, comparable in quality to careruy
.Write a note representations and mappings.
(Refer section 41)
hand-crafted natural language on
analyzers; however, they can be created for a W h a t are various approaches to krnowledge representation. (Refer section
4

language or a new domain at a small fraction of the cost in human labor"

2001, p. 8138). (Dav 3. Explain neural nets. Refer section 4.1)

Large corpora of data

may be approached by methods related 43 University Questions with Answers
which seems to be what Davis is empirrence to

however, between traditional suggesting. There is an

important u ntation Winter-14
and "text corpora" empiricist approaches to knowledge repres (Refer section 4.1.2)
approaches. In the traditional hat is
Q.1 representation.
considered knowledge by the approach is representea aplain the different issues in knowledge
person doing the representation. There
PUBLICATIONS
An up thrust for knowiedge
TECHNICAL
TECHNICAL PUBLICATIONS An up thrust fo kno