THE RELEVANCE OF THE SOCIAL SCIENCES 
TO ARTIFICIAL INTELLIGENCE AND EXPERT SYSTEMS 
' 
A study submitted in partial fulfilment 
of the requirements for the degree of 
Master of Arts in Librarianship 
at the University of Sheffield 
Mirko Popovic 
September 1986 

ABSTRACT 
This dissertation analvses the need for a closer 
association betveen the social sciences and artificial 
intelligence (Al) and expert svstems. The arguments are 
based upon a review of discussions from a variety of 
sources about Al and expert systems which reveal at 
present a wide variation in the descriptions of the 
current state of the art. 
The discussion starts with the notion of 
interpretative flexibility of the term "intelligence" 
which consequently leads to disagreements about the 
research achievements of Al. This is followed by a 
description of the main issues in Al research and its 
development which also indicates that Al is a big, and 
increasingly growing business. 
In the context of this increased interest in Al 
research and the discrepancies in reports, the section on 
sub-areas of Al and on Al as an interdisciplinary field 
reveals a narrow view of the role of the social sciences. 
On the one hand, it is often thought that the development 
of cognitive psychology is the only relationship betveen 
the social sciences and Al, and, on the other hand that 
social sciences are only concerned with the effects of Al, 
but not with its genesis. In contrast to this restricted 
view, knowledge, language, intelligence, etc., are defined 
as social concepts and the need for the social sciences' 
involvement with these main Al issues is stressed. 
This relationship is illustrated through the 
example of expert systems. Throughout a discussion about 
the main issues in expert systems, wide variations in the 
assessments of the field are presented, together with a 
indication of the oversimplified and atheoretical 
approach, mostly found in "popular" literature. This is 
also a starting-point for the analysis of the fundamental 
problems in expert systeras building, relevant to the 
social sciences, i.e. knowledge acquisition, knowledge 
representation, and explanation facilities. The main 
conclusion is' that the social sciences will only be able 
to assess the impact and effects of Al and' expert systems 
in different environments if they are also involved in 
research into these fundamental problems. 
The final chapter discusses the relevance of Al 
and expert systems research to library/in formation 
systems. New methods of organizing and representing 
information in databases , and expert intermediary systems 
are identified as two areas which could benefit from such 
research. This also has important implications for 
library/information education. 
To Breda 
CONTENTS 
ACKNOWLEDGEMENTS 
PAGE 
INTRODUCTION 1 
Chapter 1 Artificial intelligence and the role of 
the social sciences 10 
1.1. Interpretative flexibility of the term 
"intelligence" 1
1.2. Some issues in the development of Al 22 
1.3. From "General Problem Solver" to the Japanese 
Fifth Generation Computer Svstems Project, 
or: Al as a big business 27 
1.4. Sub-areas of Al 36 
1.5. Al as an interdisciplinarv field: the 
relationship between the social sciences and Al 48 
Chapter 2 Expert svstems 58 
2.1. Variations in assessments of expert svstems .... 58 
2.2. Gversimplified descriptions of main issues 
in expert svstems (a MYCIN čase study) 63 
2.2.1. Architectural principles 68 
2.2.2. Inferenče methods 71 
2.2.3. Unreliable data or knowledge 73 
2.2.4. EMYCIN as introduction to "shells" 75 
2.2.5. Explanation facilities 77 
2.3. Representation of knowledge in expert svstems .. 79 
2.3.1. Rule-based production systems 80 
2.3.2. Frame-based systems 82 
2.3.3. Logic programming systems 84 
2.4. Some features of expert systems development: 
from classic expert systems to "shells" 87 
2.5. Fundamental problems in expert systems research, 
relevant for the social sciences 94 
2.5.1. Knowledge acquisition 9
2.5.2. Knowledge representation 8 
2.5.3- Explanation facilities 101 
Chapter 3 The relevance of Al and expert systems 
research for library/information systems . . 105 
3. 1. A need for new methods of organizing and 
representing Information and knovvledge in 
databases 106 
3.2. Expert interraediary systems as front ends to 
bibliographic databases 110 
CONCLUSIONS 118 
NOTES 125 
BIBLI0GRAPHY 12
ACKNOWLEDGEMENTS 
The author would like to thank Professor M.F. 
Lynch of the Department of 'Information Studies, 
University of Sheffield, for his guidance and help 
throughout the production of this dissertation. 
Thanks are owed also to Mr. H.M. Collins, 
University of Bath; Mr. A. S. Pollitt, Huddersfield 
Polytechnic; Mr. R. Davies, Exeter University Library; 
and Mrs. L.C. Smith, University of Illinois at Urbana-
Champaign, who have supplied me with their latest papers 
and reports, relevant for my work. 
Furthermore, I would like to thank Sheffield 
University Library for its inter-library loan service 
which provided me with many valuable articles. 
Sincere thanks are offered also to my 
colleagues Amanda Saville and Carole Edwards for their 
time and effort in clarifying seme expressions in my 
dissertation. 
Finally, thanks are owed to the Culture 
Comraunity of Slovenia and the National and University 
Library in Ljubljana for their financial support. 
"... we need to investigate the relationship 
between the pronouncements of spokesmen on behalf 
of Al and the practical day-to-day activities of 
Al researchers." 
(S. Vfoolgar, 1985, p. 567) 
INTRODUCTION 
"... computerised searching services will not have 
their full impact upon user cornmunities until 
direct user searching is widespread." 
(S. Pollitt, 1986, p.1) 
There is an increasing tendency to bring coraputerized 
systems into the specialized domain of librarianship and 
information science: mechanized cataloguing and online 
information retrieval are two typical functions. 
It was apparent as early as 1968, two years 
after the MARC-project was initiated, that, 
"A machine record is not siraply a different 
physical means of recording a traditional 
bibliographic entry, for use in a traditional way" 
(Vickery, 1968, p. 1). 
Has the library world considered the arguments 
for research aimed at the determination of the structure 
and content of an optimal bibliographic record in machine-
readable form? I think not. 
Computers have been used in libraries for 
approximately twenty years. Many of the larger libraries 
have automated circulation systems, online catalogues, 
serials confcrol, financial and statistical reporting, etc. 
However, the new tools have mostly been used to perform 
the same type of work as before - witness the coraputerized 
catalogue card printer. Or, in other words, 
"... the present library systems have mostly been 
designed to the specifications of librarians, for 
librarians" (Hjerppe, 1983, p. 16). 
- 1 -
The user, his access to the text and the 
knowledge in the text and the organization of knowledge 
using the new means available, have been neglected to a 
large extent. The only "library" organi zations to 
experiment and utilize the inherent possibil ities of 
machine-readable bibliographic representation have been 
the abstracting and indexing services. 
But, however sophisticated these services raight 
be, there are many problems, most notably in online 
bibliographic information retrieval systems. Although the 
rapid retrieval of references through online systems 
normally greatly reduces the amount of tirne spent 
searching for details of documents, it has littie effect 
on other stages in the process of acquiring and using 
information. One of the greatest problems is that such 
access to knovledge does not indicate where the docunent 
carrying that information/knowledge can be found. 
The second problem in bibliographic information 
retrieval is the specification of the user's requirements 
which relates to the procedures for classifying and 
indexing the docunent. However adequate index terms or 
thesaurus entries to a docunent may be, there is one main 
difficulty: 
"The problem is caused by the nature of the 
features: the fact that they are words. Words in 
isolation can have many meanings and when combined 
into phrases they can be cajoled into many subtle 
variations that cannot be easily simulated by 
logical connectives. Consequently, the user's 
specifications to an IR system can often return 
much irrelevant material, and attempts at refining 
the original request can result in no response" 
(Addis, 1982, p. 302). 
- 2 -
Finally, untrained users can only search online 
databases having learnt artificial command languages, 
consulted manuals and relied on the help and intervention 
of trained information intermediaries. It is widely 
believed that end-users will be able to make their own 
requests online when search processes are siraplified or 
made more "friendly". 
These are same of the issues which have recently 
led to the increasing interest in library and information 
community in artificial intelligence (Al) and expert 
systems research. 
The relevance of much of this research to 
library and information service area seems obvious. Walker 
(1981) put it in these words: 
"The information retrieval systems of information 
science and the knowledge-based expert systems of 
artificial intelligence can be vieued as 
constituting two ends of a continuum of facilities 
relevant for knowledge synthesis and 
interpretation. Considered in idealized form, both 
represent static states, the content of 
information retrieval svstems providing the raw 
materials from which people derive information 
relevant for their needs; the expert svstems 
embodving digested knowledge consensually 
validated as relevant for some area of inquiry" 
(p. 360). 
One of the first attempts at bringing these two 
ends closer together can be found in Snith's article 
(1976): 
"In particular, information retrieval systems need 
no longer be limited in scope to the reference 
retrieval systems ..., but instead may be expanded 
to include fact retrieval, data retrieval, and 
question-answering as well" (p. 195). 
- 3 -
An indication of the closer association between 
Al research and 1 ibrary/information science can be found 
in the development of intermediary expert systems, the 
main aim of which is to provide aid and assistance to 
users who wish to carry out their own online searches. 
Some of these systems are already available for public 
access, and others are being tested in experimental 
settings. On the other hand, although little progress has 
been made in developing expert systems for "traditional" 
library and information work, there are more and more 
articles which stress that the use of expert systems 
forces a rethinking of the methods of organizing and 
representing knowledge and information in order to make 
them more dynamic and interactive. 
I do not intend to describe in detail the 
relevance of Al research for library/information science 
in this section - this will be done in the last chapter -
but to stress the necessity of the library/information 
profession's awareness of these developments, and to put 
the whole area into a much broader context. 
One of the main characteristics of Al research 
is the lack of a firm theoretical foundation which has its 
ground in the ill-defined term "intelligence" , and 
consequently, in endiess discussions whether machines can 
be intelligent or not. This, of course, leads to the 
discrepancies in reports about the achievements of the 
field, i.e., extraordinary optimism is countered elsewhere 
with claims that Al faces fundaraental problems. Therefore, 
- 4 -
a complex understanding of the state of the art is needed 
when applying Al research results not only to 
library/information service area, but also to ali other 
domains. 
I would like to illustrat.e this statement by the 
example of the research project "Development of scientific 
and technical information in Slovenja 1986-90" (see 
Kornhauser, 1985) which emphasizes expert systems (also 
decision-making systems) as the final step of the 
development of a whole information infrastructure. This 
project, which can also be seen as an attempt to preserve 
artificial distinction between libraries and information 
services, proposes the folloving directions of 
development, as shovm in Fig. 1.: 
Fig. 1. From information sources to expert systems ( taken 
from the research project "Development of 
scientific and technical information in Slovenja 
1986-90"). 
- 5 -
In addition, this research project does not take 
into account the different relationships between 
knowledge, communication, and information systems in the 
sciences, the social sciences, and the hunanities. It is 
evident frora Fig. 2. that this direction of development is 
equally proposed for the following complexes: 
BIOMEDICINE 
NATURAL SCIENCE 
TECHNICAL SCIENCE 
BIOTECHNICAL SCIENCE 
ECONOMICS 
SOCIAL SCIENCES 
HUMANITIES 
INTERDISCIPLINARY 
COMPLEX 
Fig. 2. Areas of application of research project 
"Development of scientific and technical 
information in Slovenja 1986-90". 
There is no doubt that the positive side of this 
project is in its introduction of "svstem-thinking" in 
information coramunitv, i.e., in linking bits of 
information into networks and shoving the 
interrelationship between data, which is proposed in the 
development from autcmated bibliographic databases, 
factographic computer-supported databases, structured 
databases, to expert systems. Some problems which demand 
new approach in organizing information and knowledge in 
databases have already been described. 
- 6 -
However, there are many questionable issues in 
this project, for exaraple: 
- uncritical adoption of "system- thinking" to ali 
disciplines; 
artificial distinction between scientific and 
technical information services and libraries; 
- unjustified reduction of libraries to the 
traditional role (libraries are even excluded from the 
automation, i.e., computerized bibliographic databases are 
only dcmain of information services) . 
Therefore, a much wider approach is needed, with 
a stress on the structure of knovrledge as a central issue. 
This is iinportant not only for library/information 
science, but also for the social sciences (especially 
sociology). In the context of this research project, there 
are many questions which should be of interest to both 
areas: 
- what are the main differences in communication 
patterns, information seeking, structure of knovrledge, 
etc., betveen sciences and social sciences, e.g., 
chemistry vs. sociology (1)?; 
how can knowledge in different disciplines be 
represented and formalized in expert systems?; 
are knowledge, language, intelligence, etc., as 
central issues of Al research, also social concepts, and 
thus not only the domain of hardvare and software 
developers and so-called "knowledge engineers"? 
- 7 -
This dissertation therefore aims to present a 
critical analysis of Al and expert systems research and a 
definition of the needs for the social sciences approach 
to Al and expert systems. The argunents below are based 
upon a review of discussions in a variety of sources about 
Al, expert systems , and their raain issues. Two raain 
characteristics of these discussions are evident: 
1 - there is a wide variation in assessments and 
descriptions of the current state of the art; 
2 - the relationship between the social sciences and 
Al research is reduced to cognitive psychology and 
linguistics; sociology is either excluded altogether or 
its contribution is only recognized in the discussions 
about the impact and effects of Al. 
The Al perspective for library/information 
systems can only be clarified by the analysis of these 
features. 
Therefore, to provide a framework for a closer 
association between the social sciences and Al research, 
and for a review of Al applications to the 
library/information service area, this dissertation begins 
with an analysis of the reasons for discrepancies in 
reports about Al. This is followed by the description of 
the raain issues in Al research and its developraent, which 
will show that Al is a big, and increasingly groving 
business. This chapter vd.ll be concluded vdth the outline 
of the main sub-areas of Al and with a discussion on Al as 
an interdisciplinary field, where the reasons for the lack 
- 8 -
of a closer relationship betueen social sciences and Al 
will be examined. 
To confirm these ideas, I will turn to examples 
taken from literature on expert systems which have been 
widely acclaiined as the applied . end of Al research. 
Throughout the analysis of the main issues in expert 
systems and the description of their development, 
disagreanents between the authors, and oversimplified and 
atheoretical approaches in the literature vri.ll be 
presented. On this basis, the fundamental problems in 
expert systeras building, which are also relevant for the 
social sciences, i.e. knowledge acquisition, knouledge 
representation, and explanation facilities, vri.ll be 
analysed. Finally, library and information systems as 
potential domains for Al and expert systems vri.ll also be 
discussed. 
- 9 -
Chapter 1 
Artificial intelligence and the role of the social 
sciences . 
"Gnce intelligence has evolved to the level of 
knowledge based on language, its social aspect 
must surely dominate" (R. Stamper, 1985, p. 172). 
1.1. Interpretative flexibility of the term "intelligence" 
The idea that the digital computer will someday 
match or exceed the intellectual abilities of human beings 
has been put forward repeatedly since its invention. In 
the past thirty years a new discipline, called "artificial 
intelligence"(2) has emerged. It is said by Waltz (1982) 
that, 
"... computer programs written by investigators in 
artificial intelligence have demonstrated 
conclusively that in certain activities (including 
activities most people vrould say require 
intelligence, such as playing games) the computer 
can outperform a human being. ... At the same tirne 
the understanding of various features of human 
intelligence has been considerably enriched by the 
attempt to describe analogues of those features in 
the detail necessary for vriting a program. As a 
result the analogy relating the performance of the 
computer to that of human intelligence has 
broadened and matured" (p. 101). 
The goals of Al research are evident from these 
statements. One of them is development of computational 
models of intelligent behaviour. A more engineering-
- 10 -
oriented goal is the development of computer programs that 
can solve problems norraally thought to require hunan 
intelligence. These are very arabitious aims, and, 
"... neither has been achieved in any general 
sense" (Duda and Shortliffe, 1983, p. 261). 
However, it can be said that few areas of 
research have been as exciting, proraising, or bevdldering 
as Al. After thirty years of use, the very name stili has 
the power to provoke controversy. 
In this context, it can also be said that Al is 
stili a relatively young science, which is characterized 
by, 
"... the lack of any clearly defined way of 
carrying out research in the field" (Ritchie and 
Hanna, 1982, p.2). 
The main problem in Al research, as shovm by Ritchie and 
Hanna (1982) on the example of AM system (i.e., system 
which has been claimed to "discover" concepts and 
conjectures in elementary mathematics) , is that published 
accounts often do not directly correspond to actual large 
and complex vrorking programs. This means that, 
"... very little of Al research fits into the 
traditional "experimental paradigm" in which well-
defined hypotheses are refuted by empirical 
investigations" (Ritchie and Hanna, 1982, p. 30). 
Although the authors in the Al community agree 
upon the lack of firm theoretical and methodological 
foundations in Al research as one of major problems, 
serious research efforts in the last ten years have led to 
important achievements and to a substantial body of 
- 11 -
fundamental principies in AI(3). There is even consensus 
among researchers about the definition of Al, as provided 
by Barr and Feigenbaun (1981): 
"Artificial intelligence is the part of computer 
science concerned with designing intelligent 
ccrnputer systems, that is, systeras that exhibit 
the characteristics we associate with intelligence 
in hunan behavior - understanding language, 
learning, reasoning, solving problems, and so on" 
(p. 3). 
What _i? controversial, quite apart from the 
subject matter, is the name itself. Conflict in Al has 
been bound up with the focus on intelligence, and it is 
written by Fleck (1982) that, 
"Intelligence is not a socially or cognitively 
well-defined goal and every distinctivs sccial 
group tends to have its own implicit definition, 
couched in teras of its own interest. Consequently 
research in Al has been oriented towards a variety 
of goals" (p. 172). 
Indeed, there are no hard and fast criteria to 
decide whether a system is artificially intelligent or 
not. Some people hold the view that intelligence is an 
essentially human attribute, and that therefore 
"artificial intelligence" is a contradiction in teras. 
Others are convinced that, however clever coraputers 
become, they will never produce anything that is genuinely 
intelligent. 
Many authors try to avoid such questions, for 
example Borko (1985), who said in one of his articles : 
"It is not my intent, nor is it necessary, to 
provide a very precise definition of artificial 
intelligence or to decide whether machines can 
think. We can leave these auestions to the 
philosophers" (p. 105). 
- 12 -
I believe, however, that understanding (not 
solving) these questions is very important from the 
follovting points of view: 
- a lively debate centring on Al reveals some 
different positions on Al research (e.g. "Al is 
impossible" vs. "Al offers a way of humanizing 
technology", etc.) (M), and consequently, leads to wide 
variations about the achievements of Al research; 
- one of the central points of the accepted definition 
of Al is that the design of intelligent systems is a 
multidisciplinary process, which can be summarized in the 
man - machine relationship. The role of the social 
sciences, which it might be expected to find the reasons 
for discrepancies in the field, is also found in this 
context. In other words, the social sciences should be 
aware of the interpretative flexibility of the term 
"intelligence", otherwise their function will be reduced 
to the analysis of the effects of Al in different 
environments. This would mean that the subject of research 
would remain in the hands of software and hardware 
specialists. 
Therefore, the heart of the problem lies in the 
question "what counts as intelligence". On the following 
pages I would like to enlarge upon this question. 
As it has been seen, the most widely accepted 
definition of Al is "designing intelligent computer 
systens, which exhibit the characteristics we associate 
vrith intelligence to human behavior". This merely imports 
- 13 -
the difficulty of using the word, in the human sense, to 
the technological sense. Michie and Johnston (1985) said 
that, 
"It is not altogether surprising that there is a 
problem with names here, since we have no sound 
definition of natural intelligence either. Some 
psychologists define it thus: "Intelligence is 
what intelligence tests measure" . So what then are 
intelligence tests?" (p. 18). 
It seems that the word "intelligence" is 
associated with endless ambiguities, or as emphasized by 
Aleksander (1984): 
"The construct is undoubtedly fuzzy, and one can 
justifiably question whether it is right to dub a 
technological area, as that of intelligent systems 
is intended to be, with this lack of precision" 
(p. 18). 
However, it is interesting to stress that, at 
the same tirne, Aleksander (1984) tries to find a solution 
in the construct of intentionality which is for him a 
shift in the paradigm of intelligent systeras, and is, 
"... very much a human construct and deals with 
our ability to relate to other people and objects, 
by understanding inwardly their likely behaviour. 
Tliis is thought to be the key construct that will 
distinguish between illusory and real intelligent 
systems" (p. 10). 
In this context, intentionality is al so used in 
the meaning • "knowing what one is talking about when 
referring to objects in the real world". 
For some years now, a major philosophical debate 
has been taking plače between leading scientists involved 
in Al on the question of whether current machines and 
programs, particularly those that process natural 
- 14 -
language, may be said to possess intentional ity. According 
to Aleksander (1984), the major proponent of the notion 
that JTO currently built machine or program has 
intentionality is the American philosopher John Searle(5). 
Searle bases his argument on the example of the programs 
the aim of which is to siraulate the human ability to 
understand stories. He introduces the problem with his 
famous Oiinese Room: 
"Suppose that I* m locked in a room and g iveri a 
large batch of Chinese writing. Suppose 
furthermore that I know no Chinese, either written 
or spoken..." (Searle, 1980, p. 418). 
He further develops his idea that this English-
speaking person, given perfect and copious memory 
facilities, could be given a mass of rules (in English) 
relating to the manipulation of Chinese svmbols. Armed 
with this, he argues, this person is in the same position 
as a language-understanding computer, and would have as 
much success in ansvering questions submitted in sequence 
of Chinese svmbols as the machine. However, no matter how 
successfully the job is performed, the performers of the 
task have no idea of the content of the story, 
"... the computer has nothing more than I have in 
the čase where I understand nothing" (Searle, 
1980, p. 418). 
Searle makes his point by showing that the 
language handling routines in a language understanding 
program are clearly just rules for handling symbols. 
Searle's major philosophical opponent is D. 
Dennett (1979), who believes that one could ase ribe 
- 15 -
intentional ity to many of the products of current work in 
/VI. His argument starts with the simple notion that 
intentionality is a concept that falls into the class of 
beliefs and is not raeasurable in quantifiable teras. 
Therefore, if I believe that a machine possesses 
intentionality, this belief could be based on a need to 
have such a belief rather than on clear-cut physical 
evidence. 
This argument is very sirailar to the one that 
the British mathematician, Alan Turing, used in the very 
early days after the invention and construction of the 
first automatic digital computer. In his paper "Ccmputing 
machinery and intelligence" , published in 1950, Turing 
addresses the question: "Can machines think?". He replaces 
an attempt to define meanings of the teras "machine" and 
"think" with a related question expressed in a relatively 
unambiguous words: "What will happen when a machine takes 
the part of A in this game?". The game referred to is the 
imitation game in which there are three participants: the 
machine (A), a human (B), and an interrogator (a machine 
and a person are in separate rooms, and both are 
interrogated by a third party via seme sort of teletype 
set-up) . The object of the game for the interrogator is to 
detennine which of the other two is human and which is the 
machine. The machine "passes" the test if the interrogator 
is unable to detennine the difference between the machine 
and the person. 
- 16 -
This experiment is commonly called "Turing* s 
test" , which, 
"... effectively says that if a human is prepared 
to attribute human qualities to an object, then he 
endows it with those qualities" (Aleksander, 1984, 
p. 23). 
1 think that it is very important to stress that 
the concept of the attribution of human qualities to 
machines, as argued by Dennett and Turing, can lead to 
seme serious over-simplifications. An illustration of this 
may be found in a program written by Bd Weizenbaum of the 
Massachusetts Institute of Technology (MIT) in 1966, 
called ELIZA. This program, which was largely designed to 
disprove Turing's notion of attribution, simulates a 
psychiatrist, encouraging the human "patient" to type in 
his troubles, in ordinary sentences, while the machine 
makes supposedly helpful cotnments in reply. A dialogue 
could look like this (6) - the cemputer's responses are 
printed entirely in upper čase: 
Men are ali alike. 
IN WHAT WAY 
They' re always bugging us about something or other. 
CAN YOU THINK OF A SPECIAL EXAMPLE 
Well, my boyfriend made me come here. 
YOUR BOYFRIEND MADE YOU CCME HERE 
He says I'm depressed much of the tirne. 
I AM S0RRY TO HEAR YOU ARE DEPRESSED 
It's true . I am unhappy. 
DO Y0U THINK COMING HERE WILL HELP YOU NOT TO BE 
UNHAPPY 
I need some help, that much seems certain. 
WHAT W0ULD IT MEAN TO Y0U IF Y0U GOT SOME HELP 
Perhaps I could learn to get along with my mother. 
TELL ME ABOUT YOUR FAMILY 
- 17 -
My raother takes čare of me. 
WHO ELSE IN YOUR FAMILY TAKES ČARE OF YOU 
My father. 
YOUR FATHER 
You are like ray father in some ways . 
WHAT RESEMBLANCE DO YOU SEE 
You are not very aggressive but I think you don' t 
want me to notice that. 
WHAT MAKE YOU THINK I AM NOT VERY AGGRESSIVE 
You don't argue with me. 
WHY DO YOU THINK I DON'T ARGUE WITH YOU 
You are afraid of me. 
DOES IT PLEASE YOU TO BELIEVE I AM AFRAID OF YOU 
My father is afraid of everybody 
WHAT ELSE COMES TO MIND WHEN Y0U THINK OF YOUR 
FATHER 
Bullies. 
DOES THAT HAVE ANTTHING TO DO WITH THE FACT THAT 
YOUR BOYFRIEND MADE Y0U CCME HERE 
and so on. 
The program is actually very simple. It looks 
for certain key phrases in the "patient's" statements, and 
for each key it has a present reply. In the above example 
the key phrases are "depressed", "my mother", etc. 
Sometimes the part of the patient1 s sentence following the 
key phrase is included in the machine's reply, with the 
pronouns and verb forms swapped round: "I" for "you" , 
"are" for "am" and so on. Several other tricks - like 
associating keywords with a class or situation ( "mother" 
implies "family") - help enhance the illusion of 
intelligent dialogue. 
- 18 -
Although ELIZA's dialogue with the user appears 
surprisingly realistic, the program does it vithout having 
the slightest understanding of the content of what it is 
repeating. If you say to ELIZA, "Let's discuss paths 
toward nuclear disarmament" , you raight well get the 
nonsensical reply, "WHY ARE Y0U TELLING ME THAT YOUR 
MOTHER MAKES PATHS T0WARD NUCLEAR DISARMAMENT" , if you had 
introduced the word "mother" in your previous interchange. 
It is said by Michie and Johnston (1985) that, 
"ELIZA is nothing but a very carefully worked-out 
parlour trick. Vfeizenbaum intended it as a joke -
a parody - and was appalled when established 
psychiatrists took it seriously and started 
talking about the possibility of automated 
psychotherapy" (p.25). 
And indeed, raanv who encountered ELIZA 
attributed human properties of understanding, and even 
interest, to the simple canputer program. Weizenbaum was 
shocked at the misinterpretation of his work and noticed 
three distinct results: 
1 - a number of practicing psychiatrists seriously 
believed that such canputer programs could grovr into a 
nearly completely automatic form of psychotherapy (7); 
2 - seme people, conversing with ELIZA became 
emotionally involved vdth the computer; 
3 - there vas also a spread of belief that ELIZA 
demonstrated a general solution to the problem of computer 
understanding of natural language. 
- 19 -
These conclusions have led Vteizenbaum to discuss 
the dangers of work in the field, and his condemnation of 
excessive faith in technology can be found in his book 
"Computer power and human reason", first published in 
1974. 
I think that the above examples have showed that 
the quest for intelligent machines cannot rest only with 
the attribution argument. But why have I discussed the 
problem of intelligent machines without giving any final 
solution? 
There is no doubt that these polemics among Al 
researchers reveal the interpretative flexibility 
associated with the notion of intelligence and 
intentionality. This feature begins to account for the 
variations in reports of the state of Al research, which 
can be found particularly in the example of expert 
systems, where, 
"... we might expect optimistic representations of 
the vitality, achievements and potential of the 
field from those involved in marketing expert 
systems" (Woolgar, 1985, p. 564-565). 
And indeed, expert systems research is a field 
vrhere the extraordinary optimism of seme reports is 
elsewhere countered by considerable caution and pessimism 
about the achievements to date. On the one hand, expert 
systems are generally regarded as one of the most active 
areas of Al research, and on the other hand, there is 
considerable concern about the fact that the field 
currently faces fundamental problems (see, for example, 
Dud a and Shortliffe, 1983; Leith, 1986). 
- 20 -
What are the impl ications of the features of Al 
discourse outlined above? To recognize the interpretative 
flexibility of notions of " intelligence" is very important 
for the social sciences because, 
"... adherence to the view that the phenoraenon for 
Al investigation are the inner processes 
responsible for "thought" and "intelligence" vd.ll 
plače these entities beyond the reach of mere 
observational social sciences" (Woolgar, 1985, 
p.565) 
Indeed, as it vd.ll be illustrated later, the 
role of the social sciences among Al researchers is too 
often seen only in investigations of the impact and 
effects of Al, or the input of social sciences in Al 
research is too often reduced on cognitive psychology. At 
this point, the main idea can be emphasized: the social 
sciences vd.ll be able to assess the impact of Al only if 
they al so become involved in a detailed consideration of 
the processes of research activity in Al. Therefore, the 
social sciences should also be concerned with the genesis 
of Al, and not only vdth its effects. In this context I 
vrould like again to cite Aleksander (1984) who says, 
"... the crucial issue in assigning any form of 
hunan wisdom to a machine is that we must be able 
to understand plainly how this vdsdom gets into 
the machine in the first plače" (p. 25). 
Wisdom - known in Al community under name 
knowledge is the central issue. To understand how human 
knov/ledge can be encapsulated (or encoded) into Al 
applications (e.g. expert systems) is the starting-point 
for a discussion about the impact and effects of Al. These 
- 21 -
discussions (e.g., Al and unemployment - see Partridge, 
1986; how the status of the human expert will be affected 
by expert systems, legal implications of the use of expert 
systems, etc. - see Boden, 1984) cannot be realistic 
vdthout this notion. 
These ideas will be followed throughout my 
dissertation, with the main eraphasis on knouledge 
representation, which will be analysed in the context of 
expert systems. But first, I would like to clarify the 
main issues in Al research and to give a short description 
of development of Al with the notion of Al as an 
increasingly growing business. On this basis, the problem 
of Al as an interdisciplinary field vd.ll be discussed, 
where the need for the closer association between the 
social sciences and Al will be stressed. 
1.2. Seme issues in the development of Al 
Al had its origins in the late 1950s and early 
1960s when it was recognized that electronic computers 
were more than giant calculators: they could process 
symbols, expressed as nunbers, letters of the alphabet, or 
words in a language. 
The second impulse for Al research can be found, 
aceording to Dud a and Shortliffe (1 983), in the shifting 
goal of much science which originally tried to obtain only 
quantitative deseriptions of natural phenomena. 
- 22 -
Unfortunately, not ali natural phenomena can be expressed 
well in numbers. In particular, symbolic rather than 
nuraerical operations seem to characterize such activities 
as planning, problem-solving and deduction. Serious work 
on Al began when it was real ized that computers as 
processors of symbols are potentially capable of being 
programmed to exhibit such intelligent behaviour. This 
non-numerical emphasis is a crucial characteristic of Al 
which distinguishes Al from the mainstream of computer 
science. 
According to Newell (1983), there are two more 
factors that served to isolate Al within computer science: 
1 - its choice of heuristic programming techniques, as 
distinct from algorithms favoured by computer scientists; 
2 - its development of list-processing program 
languages, when the rest of computer science was moving 
toward the use of ccmpilers. 
What is a difference between heuristic programs 
and algorithmic programs? An algorithm is a precisely 
defined procedure consisting of a series of steps or 
program instructions for performing a specific task which 
would necessarily lead to a problem solution. In contrast, 
heuristic programs utilize approximate and exploratory 
methods based upon partial knowledge vihich might lead to 
the discovery of a problem solution but which could not be 
guaranteed to do so. Heuristics enable one to vrark with 
ill-defined problems. A classic example from Al research 
is the game of chess for which no algorithmic solution 
- 23 -
exists; heuristic rules could be programmed, for example, 
to choose between two moves by selecting the one that 
restricts the opponent's raobility to the larger degree. 
Heuristic programs incorporate procedures for selecting 
alternativ es and evaluating the results of partial 
solutions while progressing toward a final goal. 
Lis t-proces s ing languages, for example LISP (8), 
are ccmputer languages that facilitate the processing of 
data organized in the form of lists. They transform a 
program statement into a sequence of machine actions. In 
contrast to these are compiler languages (e.g., COBOL, 
FORTRAN, PASCAL, etc.) which were originally used for 
numerical computations, and are transformed directly into 
machine language. Canpiled programs can be executed with 
much greater speed, but list-processing languages allow a 
higher degree of interaction and user involvement, a 
matter which is one of the major concerns of Al 
researchers. 
With regard to these issues, the initial work in 
Al limited itself to non-numeric but well-defined and 
well-constrained problems such as symbolic algebra, chess 
playing or game playing in general, puzzle solving, and 
simple theorem proving. According to Nilsson (1971), among 
the important techniques that emerged were general methods 
for representing information in syrabolic data structures, 
general methods for manipulating these structures, and 
heuristics for searching through them. 
- 24 -
But, as it is stressed by Sowizral (1985), 
".. . most of the early work in artificial 
intelligence served only to tease researchers. It 
hinted at the feasibility of canputer reasoning 
but fell far short of solving practical problems. 
Clearly, problem- independent solution methods 
could not handle the combinatorial complexity of 
real-world problems" (p. 180). 
Therefore, Al researchers started asking how 
people solved real-world problems. A frequent answer was 
that people possess knowledge of which the programs vere 
wholly innocent. Or, as it is said by Lenat (1984): 
"By the mid-1970's, after two decades of 
humblengly slov/ progress, workers in the new field 
of artificial intelligence had come to a 
fundamental conclusion about intelligent behaviour 
in general: it requires a tremendous amount of 
knowledge, which people often take for granted but 
which must be spoon-fed to a computer" (p. 152). 
The central role of knowledge in intelligence 
explains why the most successful programs so far have been 
expert systems which operate in highly specialized 
domains. But to be efficient, this knov/ledge has to be 
ccmbined with methods of conceptualizing and reasoning 
about the problem area. Or, in other words, 
"... in attacking a ccmplex problem people draw on 
various methods - I call thera sources of power -
of using their knovrledge of the world' s 
regularities to constrain the search for a 
solution. They may invoke mathematical theorems or 
less formal rules of thumb; they may break up the 
problem into more tractable subproblems, or they 
may reason by analogy to problems that have 
already been solved. To the extent that computer 
programs already exhibit intelligence it is 
because they draw on seme of these same sources of 
power. The future of artificial intelligence lies 
in finding ways to tap those sources that have 
only begun to be exploited" (Lenat, 1984, p. 152). 
- 25 -
The groving recognition of the many kinds of 
knowledge required for high-performance reasoning systems 
changed the shape of Al research. Not surprisingly, 
research began shifting from the development of powerful 
and general but combinatorially_ expensive reasoning 
techniques to the developrnent of effective techniques for 
representing large araounts of knowledge and effectively 
using that knowledge. In the vrords of Goldstein and Papert 
(1977), 
"Today there has been a shift in paradigm. The 
fundamental problem of understanding intelligence 
is not the identification of a few powerful 
techniques, but rather the question of how to 
represent large amounts of knowledge in a fashion 
that permits their effective use and interaction. 
... The current point of view is that the problem 
solver (whether man or machine) must know 
explicitly how to use its knovledge - with general 
techniques supplemented by dcraain-specific 
pragmatic know-how. Thus, we see Al as having 
shifted from a power-based strategy for achieving 
intelligence to a knowledge-based approach" (p. 
84). ~~~" ' 
The result of this shift was a rapid growth of 
Al research and its applications in a nunber of fields. 
The most successful programs so far, as I have already 
said, have been eocpert systems, development of which can 
be seen in the follovdng rhetorical assertions: "The 
science of artificial intelligence ... is at last emerging 
from academic obscurity" (Evanczuk and Manuel, 1983, p. 
139); "Expert systems provide 'practical ušes for a 
useless science'" (Alexander, 1982, p.1); "Knowledge-based 
expert systems cone of age" (Duda and Gaschnik, 1981, p. 
238), etc. 
- 26 -
Although expert systems vd.ll be analvsed in a 
separate chapter, it is necessary at this point, because 
of their commercial implications, to illustrate the shift 
in Al paradigm by the example of the increasing interest 
in Al research, as expressed in the Japanese Fifth 
Generation Ccmputer Systems Project, Alvey Programme in 
the U. K. , etc . 
1. 3- From "General Problem So Iver" to the Japanese Fifth 
Generation Computer Systems Project, or: Al as a big 
business 
One of the characteristics of today's Al 
research is that the central core of research tools is 
applied in a bevri-ldering and increasing variety of 
application areas - language understanding, expert 
svstems, vision and robotics, to mention but a few. The 
multipiication of research areas - which will be described 
in one of the coraing sections - accompanies and reflects a 
move from the great optimism in the early days of Al 
research during the 1950s and early 1960s, to a period of 
stagnation in the early 1970s, and to an explosive growth 
of interest in Al in the early 1980s. 
Early optimism vas evident, for exaraple, in the 
attempts to build a "General Problem So Iver" (see Ernst 
and Newell, 1969) which could deal vri.th any area of 
knovledge. After many years of effort, refiecting the 
- 27 -
extent and complexity of the areas of knowledge concerned , 
this approach proved fruitless and the goal of a general 
intelligent inference mechanism vas abandoned . 
In the early 1970's Al research had been forced 
into the background. The situation was especially 
interesting in the U.K., where, according to Manchester 
(1986), the Lighthill report (under the auspices of the 
Science Research Council) halted Al research by 
recommending that governraent funding should be curtailed. 
The report said that the lack of a bridging 
technology between theoretical, ccmputer-based research in 
automation and research into neurobiology and psychology 
was a barrier to progress. Lighthill wanted results and to 
bring research to the point where it could generate 
scmething resembling a conmercial product. Otherviise, the 
report noted, there vere, 
"... doubts about vhether the vhole concept of 
artificial intelligence as an integrated field for 
research is a valid one" (cf. Manchester, 1986, p. 
38). 
As a result of this, Al research in the 1970s 
vas continued in a much lover key, vith more realistic 
goal s, but vith deeper ideas and methods and vith better 
programming tools. One of the signs of raaturity vas the 
already mentioned recognition that knovledge is as 
important as reasoning. 
In the beginning of the 1980s, Al vas suddenly 
again in the centre of attention, not only in the academic 
vorld and ccmputer corporations but, significantly, 
- 28 -
national governments, especially in the USA, UK, and 
Western Europe. There are two main reasons for this 
tremendous turning-point: 
1 - successful development of expert svstems and their 
commercial potential; 
2 - launching of the Japanese Fifth Generation 
Computer Svstems Project which was based on the estimation 
that commercial success lies in the use of Al methods and 
tools for the fifth-generation computers. 
The fact is that the Japanese Fifth Generation 
initiative, at an international conference in Tokyo in the 
Autumn 1981, helped to crvstallize an interest which had 
alreadv been growing in the West, but placed the 
development targets well bevond those that most Western 
researchers vrould have set for themselves. The report of 
the Japan Information Processing Development Center - the 
Jipdec Report - stated the target in these words: 
"The Fifth Generation Computer Svstems will be 
knowledge information processing svstems having 
problem-solving functions of a very high level. In 
these systems, intelligence will be greatly 
improved to approach that of a human being" (cf. 
Bramer, 1985, p.3)• 
In their well-known book, Feigenbaum and 
McCorduck (1984) added that, 
"... the Japanese expect these machines to change 
their lives - and everyone else's. ... Their Fifth 
Generation plans say unequivocally that the 
Japanese are the first nation to ... have acted on 
a truth that has been emerging and reiterated for 
nearly two decades. The world is entering a new 
period. The wealth of nations, which depended upon 
land, labor, and capital during its agricultural 
and industrial phases ... will corae in the future 
to depend upon information, knowledge, and 
intelligence"(p. 14). 
- 29 -
But the concept of the fifth-generation computer 
systems, as illustrated by the widely circulated figure on 
the next page, i .e. Fig. 3., presupposes an implementation 
of Al-based subsysteras with capabilities beyond those 
feasible in research today. However, this project is a 
good exaraple of what has been considered to be within 
reach in teras of the application of Al-based concepts. 
There is no doubt that many of the themes for fifth 
generation computer systems arise from prior research in 
Al. While this diagram may seem confusing, it clearly 
shows the determination to move in the direction of 
intelligent knowledge-based systems. The Japanese are also 
talking about the main programming of their machines being 
carried out in a "logic programming language", a technique 
originating from Al work and differing radically from 
conventional computer languages. And this softvare should 
run on fifth generation hardvare which vdll feature 
parallel processing in contrast to current sequential 
machine architectures. The interest of the Japanese 
government to support such research can be illustrated, 
according to Clarke and Cronin (1983), with the ?200 
raillion which were provided for this ten-year project. 
It has already been said that the Japanese 
project sounded the alarm in the Western computer 
industry. One of responses was the British Government's 
Alvey Programme (9) for Advanced Information Technology. 
With the report of the Alvey Committee in 1932, Al was 
"rehabilitated" in official circles in the U.K., under the 
- 30 -

new name of IKBS (Intelligent Knowledge-Based Systems) . It 
is amusing to note the comment in t ne Alvey Report, after 
the highly damaging Lighthill Report, that, 
"The need to train additional personnel ... is 
particularly pressing in the IKBS area, where 
there are at present few active. participants" (cf. 
Braraer, 1985, p. 1). 
The Alvey Programme (10) is defined as a joint 
venture between three UK Government Departments - the 
Department of Trade & Industry, the Ministry of Defence, 
and the Department of Education and Science - and British 
Industry and academia. It is a five-year programme, begun 
in 1983 and costing ?350m, of which ?200 mili ion coraes 
from public funds and ?150 million from industry. Its 
objective is to stimulate British information technology 
research through a programme of the following projects: 
- Intelligent Knowledge Based Systems (IKBS); 
- The Man-Machine Interface (MMI); 
- Software Ehgineering; 
- Very Large Scale Integration (VLSI); 
- Computing Architectures. 
The main interest for us is, of course, IKBS 
which is divided in to the follovdng four sub-programmes: 
- IKBS Demonstrators; 
- Research Theraes, Projects and Club s; 
- Support Infrastructure; 
- IKBS Awareness. 
The demonstrator projects are intended to apply 
the ideas and techniques developed by the research 
community in systems viewed by industry as the precursors 
- 32 -
of market products. There are four IKBS demonstrator 
projects, ali featuring the use of IKBS techniques in some 
specific area of industry, i.e., raechanical engineering, 
cheniical industry, systems engineering, and applications 
in the data processing industry. 
Ali research work in IKBS is organized within 
the framevrark of research themes and clubs, which have a 
cotnmon policy on such matters as tools and standards for 
languages. The structure for the year 1985, together with 
its financial implications is shown in Table 1. 
Name of club 
Knouledge 
Based Systems 
Logic Based 
Environments 
Declarative 
Architectures 
Speech and 
Natural 
Language 
(jointly with 
MMI) 
Vision 
(jointly with 
MMI) 
No of 
projects 
24 
6 
19 
8 
H 
Cost 
(? mil.) 
6.4 
0.5 
16.7 
0.8 
1.8 
Research theraes 
IKB Demonstrator s 
Large demonstrator 
projects (IKBS 
components) 
Expert systeras 
Intelligent front 
ends 
Intelligent coraputer-
aided instruction 
Declarative languages 
Inference and 
knowledge 
representation 
Parallel architectures 
Intelligent database 
systems 
Natural language 
Image interpretation 
Table 1. Research themes, projects and clubs in the IKBS 
for the year 1985 (after Alvey Programme, Annual 
Report 1985) 
- 33 -
The need to bring to the attention of a wide 
speč trum of UK organizations the potential future 
importance of IKBS techniques, in general, and expert 
systeras in particular, also led to the setting up of an 
IKBS awareness program and to the formation of several 
industry specific expert systems clubs. Typically, each 
club has 20 industrial organizations as members and the 
funds accumulated in this way are used to commission the 
building of expert systems in the area of interest to the 
club. At this moment, these clubs cover the following 
areas: real-tirne process control, insurance, transport 
industry, econcmetric modelling, data processing, and 
computer system fault diagnosis. 
Another response to the Japanese Fifth 
Generation Computer Systems Project came in 1982 from the 
Commission of the European Communities which stressed the 
need for a European Strateg ic Research Programme in 
Information Technology - ESPRIT - based on the argument 
that information technologies represent the fastest 
growing sector of industry today. 
Finally, as an indicator of the grovdng interest 
in Al research in the US this table is presented taken 
from Hayes-Roth (1985a) which estimates the levels and 
rates of change of seme key technology measures : 
- 34 -
Item 
Knowledge system prototypes under 
developnent 
Knowledge systems being deployed 
Knowledge systeras being raaintained 
Knowledge engineering departments 
established 
Senior knowledge engineers 
Knowledge engineers 
Knowledge engineer trainees 
Applied Al graduate students 
Undergraduate students in Al 
LISP or PROLOG programmers 
LISP or PROLOG installations 
1984 
level 
70 
15 
10 
15 
40 
150 
300 
250 
2000 
2000 
400 
1984-1985 
change 
+50% 
+100% 
+200% 
+150% 
+50% 
+50% 
+100% 
+20% 
+50% 
+50% 
+100% 
Table 2. Estimated measures of current US technology 
capacities (after Hayes-Roth, 1985a, p. 22) 
The greatest benefit of the explosive growth of 
interest in and work on Al and expert systeras can be seen 
from these examples, i.e., the much greater openness of 
commerce and industry to the ideas, technicjues and tools 
of Al and the far greater villingness to experiment with 
building systems of their own. According to some of the 
latest studies (see Manchester, 1986, p. 38) the vrorld 
market for • Al products reached $342 million in 1985 
ccmpared with $181 million in 1984. These studies forecast 
that the market will be worth $665 million by next year. 
About half of that figure will be spent on software and 
35% of the software market will be for expert systems 
products. On the bas is of these figures it can be 
concluded that Al is a big, and increasingly growing 
business, at present most notably expressed in the fields 
of expert systems, natural language, and robotics. 
- 35 -
As a result of this grovdng interest in Al it is 
necessary in the context of this study to give a short 
description of raain sub-areas of Al research. 
1.4. Sub-areas of the Al 
When trying to outline the main Al sub-fields, 
there is again a problem of different views, and 
consequently a lack of firm theoretical foundations for 
Al. On the one hand, Aleksander (1984) identifies four 
major areas of Al, i .e., game playing, problem solving, 
artificial vision, and natural language understanding. On 
the other hand, in Fleck's article (1984), thirteen sub-
areas of Al can be found. 
To follow the main aim of this section, i.e., to 
outline some characteristics of Al sub-fields, help can, 
perhaps, be found in the so-called "Al-pie" - but not as 
the ultimate valid approach - as presented by Cercone and 
McCalla (1984) and shown on the next page (see Fig. 4.). 
As it can be seen from Fig. 4. , major ef forts 
into Al research have concentrated on natural language 
understanding, computer vision, learning, theorem proving 
and logic programming, search, problem solving and 
planning, expert systems, knowledge representation, and 
other categories such as intelligent computer-aided 
instruction and tutoring, game playing, speech, automatic 
programming, and Al tools. At this point it is interesting 
- 36 -
to note that at present greatest attention in the Al 
community is paid to expert systeras and natural languages. 
According to Snith (1985), papers submitted for the 1985 
International Joint Conference on Al included 111 on 
expert systems, 99 on natural language understanding, and, 
interestingly for later discussion, only 4 on social 
implications. 
Fig. 4. The "AI-pie" (after Cercone and McCalla, 1984, p 
281). 
In the fcllcving sub-seccions each sub-area in 
the "Al-pie" will briefly be outlined. 
- 37 -
1. Natural ianguage understanding: this has been one 
of the major research areas since the earliest days of Al. 
In the 1960s raachine translation projects dcminated , but 
they failed to account for meaning, context, etc. Winograd 
(1972) was the first to suggest a solution, by noting that 
conversations have to be about something. He suggested 
that the conversation should be about a restricted world. 
Further, he proposed that the extraction of meaning may be 
guided by a process of grammatical parsing. 
According to Cercone and McCalla (1984), 
understanding natural Ianguage involves three levels of 
interpretation: 
1 - syntactic processes "parse" sentences to make the 
grammatical relationships betueen words in sentences 
clear; 
2 - semantics is concerned vdth assigning meaning to 
the various syntatic constituents; 
3 - pragmatics attempts to relate individual sentences 
to one another and to the surrounding context. 
The boundaries separating these levels are not 
distinct. At this moment the follovdng directions of 
research into ali levels of natural Ianguage exist: 
- exploring alternative powerful parsing techniques; 
- developing various schemes for explaining the 
semantics of natural Ianguage; 
- modelling connected discourse and dialogue, 
especially focussing on pragmatic issues such as story 
structure, fccus, reference, etc.; 
- 38 -
- building natural ianguage systeras, e.g., front-ends 
to database systems. 
One of the serious limitations of present-day 
natural-language systems is that they only work vdthin a 
very limited domain of discourse. Their main advantage is 
that they enable a user to interact with databases without 
the use of specialized machine programs, for example so-
called question-answering systems (e.g., LUNAR, developed 
in the early 1970's to allow lunar geologists to 
conveniently access, compare and evaluate the chemical 
analyses on lunar rocks and soil compositions accumulated 
by NASA during the APOLLO programme; and the LADDER 
system, which has been developed at SRI International and 
operates on a large naval database). 
2. Computer vision: this is another very active, and 
very difficult area of Al research. Its basic objective is 
to interpret pictures (rather than to generate pictures 
which preoccupies computer graphics). Much research has 
been done into the problem of "pattern recognition", some 
of it with computers trving to make sense of television 
images of scenes consisting of simple geometrical objects: 
blocks, pyramids, boxes, etc. Sometimes the computer 
manipulates the object with a robot arm. Among the things 
the machine has to understand are that: the view of an 
object can be obscured by another in front of it; every 
thing must be supported by something or it will fall, etc. 
- 39 -
Central to the problem is the fact that a 
picture contains an enorraous araount of infonnation. 
According to Michie and Johnston (1985), it is out of the 
question to do this processing with a conventional 
coraputer, because essential to the principle of such a 
raachine is that it processes everything in a strict 
sequence, one item of data after another, and present 
circuits just cannot move fast enough. Ovring to this "von 
Neumann bottleneck" , for image processing in particular, 
infonnation must unavoidably be processed many bits at one 
tirne, that is, in parallel. For machines to do this, a 
completely new type of hardware is needed. 
Seme machine vision systems for robots have, 
hovrever, already reached the stage of being marketed (11). 
A very interesting situation emerges, if robotics is 
discussed from the commercial point of view. At the 
beginning of their development, there were great hopes of 
robots being general purpose machines. In the event, those 
hopes were soundly dashed. According to Fleck (1984), 
there vere fewer than 200 units in use in the whole of the 
UK at the end of 1979, and, in general, diffusion 
everywhere was much slower than the manufacturers and 
premoters had expected, with only some 20-30 thousand 
robots in use worldwide by 1983. Practical experience has 
clearly deraonstrated that certain robots are best suited 
to particular tasks within a narrow range of applications. 
At present, a differentiated set of more articulated and 
specific aims, with specialized knowledge and expertise 
- 40 -
developing around them, structures and guides research and 
developraent. 
In fact, so far, industrial robots have been 
slow to diffuse, their economic feasibility has been 
difficult to demonstrate, and robot manufacturers have 
found it hard to achieve profitability. Despite this, 
"... excitement stili prevails and there is rauch 
activity, with well over two hundred manufacturers 
in what is a relatively small market. ... This 
interest is based on the assumption that robots 
will be of great importance in the future. ... 
Robots have become a symbol of national 
technological progress, a sort of international 
virility svmbol, to such an extent that many 
companies have already introduced them vdthout 
concern for the economics, to prove to themselves 
and others that they can handle new technologyri 
(Fleck, 1984, p. 208). 
3. Search, problem solving, planning: it has already 
been said that the first big "successes" in Al were 
programs that could solve puzzles and play games like 
chess. Techniques such as looking ahead several moves and 
dividing difficult problems into easier sub-problems 
evolved into the fundamental Al techniques of search and 
problem reduction. 
In general, there are three main problem-solving 
technigues: 
- state-space search, vrhich is nicely described by 
Nilsson (1982) through the example of the 15-puzzle. The 
main idea behind this kind of search is that we need to 
find a path from some initial state to any (one or more) 
goal state by applying operators to transform states into 
-41-
other states. This is also called forvrard direction of 
searching. State spaces can also be searched in a backvrard 
direction by starting vdth the goal state and applying the 
inverses of the operators to find a path to the initial 
state. Which approach is more appropriate depends on the 
particular problem and the nature of the state space; 
- propagation of constraints: in this technique, the 
set of possible solutions becomes further and further 
constrained by rules or operators that produce "local 
constraints" on what small pieces of the solution must 
look like. More and more rule applications are made until 
no more rules are applicable and only one (or some other 
small number) possible solution is left. This process can 
be thought of as a type of state-space search that avoids 
the necessity of backtracking, since every existing 
solution must satisfy ali the constraints produced by the 
rule applications; 
- problem reduction: in this technique, the problem to 
be solved is partitioned or decomposed into sub-problems 
that can be solved separately, in such a way that 
combining the solutions to the sub-problems will yield a 
solution to the original problem. Each sub-problem can be 
further reduced, until "primitive" probleras, which can be 
solved directly, are generated. Some deccmpositions of a 
problem may lead to solvable sub-problems, others may not. 
Problem solution is represented by a solution graph. 
A typical example of an expert system which is 
based on problem reduction in conjuction with forvrard 
-42 -
searching is DENDRAL (i.e., a coraputer svstem that 
proposes plausible cheraical structures for molecules, 
given their mass spectrograras) , which ušes rules to narrov/ 
the searches to manageable nurabers. This approach is 
called "reasoning by eliminating", -and is based on early 
pruning, as illustrated in Fig. 5.: 
Fig. 5. "Pruning" a search tree. The shears indicate a 
plače where the svstem could have grown a whole 
extra sub-tree in its search, but vas saved the 
labour by the intervention of a pruning criterion 
which indicated lack of prcraise in that direction 
(after Michie and Johnston, 1985, p. 44). 
An alternative approach, when the coraplexity of 
problem-so Iv ing methods increases, is producing plan s . 
- 43 -
Planning is preparing a program of actions to be carried 
out to achieve goals. An example is experimental planning 
in molecular genetics (see Stefic, 1981). A planner is 
required to construct a plan that achieves goals vdthout 
consuraing excessive resources or violating constraints. 
The techniques described above underlie most 
areas of Al and are also used in expert systems building. 
4. Expert systems: they are one of the most active and 
exeiting areas of applied research in Al. Expert systeras 
use Al problem solving and knovrledge representation 
techniques to combine human expert knowledge about a 
specific problem area with human expert methods of 
conceptualizing and reasoning about that problem area. 
Because of the central role of expert systems in 
my study they will be separately analysed in the follovang 
chapter. At this stage it can only be said that the work 
of the last few years has shovm that programs which can 
operate at or near the level of human experts are 
feasible; several have been demonstrated as being capable 
of such performance in carefully selected, well-specified 
domains. As a result, the field is beginning to undergo a 
transition from basic research to applications, which has 
resulted in increasing commercial and industrial interest. 
5. Theorem proving and logic jprogramming: this area 
has also been significant in the development of Al 
- 44 -
research. Theorem proving refers to the process of making 
logical deductions starting from a noncontradictory set of 
axioras specified in predicate calculus (firsts order 
logic). Robinson (1965) showed how it was possible to 
totally automate this process using a raethod called 
resolution. Theorem proving is al so at the heart of the 
more recent development of the programming language PROLOG 
(12). 
6. Knowledge representation: the central role of 
knowledge in building "intelligent machines" has already 
been stressed and there is no doubt that consequently, 
"... the representation of knowledge is the key 
issue in the development of Al" (Barr and 
Feigenbaura, 1981, p. 59). 
But, surprisingly, although many representation 
methods have been developed in the last thirty years, the 
most important being logic, semantic nets, production 
systems, and frames, there is stili no consensus on this 
topic. Many survevs of knowledge representation reveal a 
large variety of different views. Because of their 
importance, knowledge representation schemes will be 
outlined in the context of expert systems. 
7. Learning: there are two very well known programs in 
"learning community", i.e., AM, and EURISCO. Lenat (1977) 
constructed a program (AM) that used heuristic search 
techniques to "discover" (although not prove) new concepts 
and theorems in mathematics from about hundred elementary 
- 45 -
concepts in set theory. A follow-up program, EURISCO 
(Lenat, 1982), showed that similar methods could work in a 
uide variety of domains (e.g., fleet design, VLSI design, 
e te .). 
This is a very iraportant area of Al research 
because unless a coraputer can expand its own capabilities 
on the basis of "experience", the performance of the 
program is limited by the knowledge, foresight and 
available tirne of the programmer. 
8. Other areas; there are a nunber of other areas that 
are often included in categorizations of Al research, 
including: 
a) computerized game playing: interest in automating 
the game playing process has been manifested in Al since 
its inception as a field, not only for the obvious 
interest of getting a computer to play games well, but 
also because it was thought that the lessons learned by 
programming game playing strategies vrould generalize to 
the rest of Al. Hovever, more recently, as Al programs 
have become more "knovdedge intensive" , the so-called 
"weak, general" methods used in game playing have beceme 
less and less relevant. Although current game playing 
programs are extremely competent (e.g., there are game 
playing programs today that play near-master level chess) , 
game playing as a research area is now pursued more for 
its intrinsic interest than for the lessons it can give to 
other areas of Al. 
- 46 -
b) Al applications in education: there have been two 
main directions: 
- producing intelligent tutoring systeras that can 
behave with more subtlety and knowledge than traditional 
computer assisted instruction systems; 
- developing learning environments for students, 
e.g., the LOGO programming system frora which students can 
learn about prograraming and gecmetry. 
c) Al tools: this field consists of developing 
programming languages (LISP, PROLOG), knovledge 
representation languages, and also of building specialized 
hardware (e.g., LISP machines). This area is of great 
commercial potential. 
d) a number of other areas are also identified with 
Al, for example: speech understanding, automatic 
programming, etc. 
I hope that this revievr of the main sub-areas of 
Al has provided some ideas about the canprehensiveness of 
the field. In short, research in Al can be characterized 
by the kind of activity or area of behaviour studied, or 
by the basic concepts and techniaues that reflect the 
underlying mechanisms. In the first čase, it is 
appropriate to refer to computer vision and robotics, 
language understanding, expert systems, etc. Considered in 
relation to the concepts and techniques, Al is concerned 
with issues of knouledge representation, search and 
problem solving procedures, logic programming, etc. 
- 47 -
It can be conciuded that this discussion has so 
far revealed scrae interesting features of Al research. One 
of them is that Al is not more the "property" of the 
academic vrorld, but is becoming an iraportant ccmmercial 
field. As a result of the growing interest in Al, there 
are raany calls in the Al coramunity, addressed to the 
social sciences for investigations into the impact and 
effects of Al (see Daviš, 1982; Boden, 1984). 
At this point, I would like to emphasise the 
main argument once more: the social sciences will only be 
capable of assessing the impact of Al if they also have 
insight into key issues of Al research, for example, 
representation of knowledge. To testify this notion, it is 
necessary to open a discussion about Al as an 
interdisciplinary field and its relation to social 
sciences. 
1.5. Al as an interdisciplinary field: the relationship 
between the social sciences and Al 
Although there are again many difficulties in 
trying to define the multiplicity of Al roots, it seems 
that there is at least one consensus among Al researdhers: 
Al is a part, although an isolated part, of ccmputer 
science. The problem arises when trying to clarify the 
association between other disciplines and Al research. 
Minsky (1979), one cf the pioneering researchers in this 
- 48 -
field, suggests that Al shares its goals with the 
follovang disciplines: 
"With coraputer science we try to understand ways 
in which in formation-using processes act and 
interact. With philosophy we share problems about 
mind, thought, reason, and feeling. With 
linguistics we are concerned with relations among 
objects, symbols, words, and meanings. And with 
psychology we have to deal not only with 
perception, memory, and such matters but also with 
theories of ego structure and personality 
coherence" (p. 400). 
It can be said that this is a "classic" view of 
the interdisciplinarity of Al, shared among the majority 
of Al researchers which stresses the central role of 
computer science, philosophy, linguistics, and psychology. 
In the context of the notion that knowledge 
representation is a central issue of Al research, the main 
interest lies in an analysis of the relationship betveen 
Al and the social sciences. 
For many years it has been thought that the 
development of cognitive psychology was the only 
relationship between the social sciences and Al. The main 
idea behind this association vas based on the argument 
about the validity of so-called "strong programme of Al". 
This programme relies on the adequacy of the 
"computational metaphor": a belief that the human mind can 
be studied as though it were a computer. For example, it 
is presumed that understanding speech involves 
computational processes in the brain that are similar to 
the processes performed by an Al program designed to 
accept natural language. It has been said by Gilbert and 
- 49 -
Heath (1985) that, 
"The computational raetaphor iraraediately suggests 
that Al and cognitive psychology have much in 
coraraon, and indeed this has been for many years 
the perspective of the majority of cognitive 
scientists and Al practitioners" (p. 2). 
There is ne doubt that Al research had a 
significant influence on the birth of cognitive psychology 
and thereby cognitive science during the 1970s. The Al 
concern with "intelligence" makes its close links with 
psychology unsurprising. For example, in her well-known 
book, Boden (1977) argues that Al, 
"... offers an illuninating theoretical metaphor 
for the mind that allows psychological questions 
to be posed with greater clarity than before" (p. 
473). 
However, as Al researchers have begun to tackle 
the difficult problems in understanding natural language, 
in representing knouledge and belief, in planning actions, 
and other areas, they have looked around at other 
disciplines to see how they have approached these issues. 
For instance, Al has been influenced by, and has in turn 
itself influenced, linguistics, 
The association, therefore, betveen Al, on the 
one hand, and disciplines like psychology and linguistics 
on the other hand has been widely recognized and debated 
in teras of the implications of one for the other. But, 
sociology, 
"... whose interests clearly eneorapass language 
use and interaetion, belief and knovdedge systetns, 
aetion and intentionality, is as yet unexplored 
territory" (Gilbert and Heath, 1935, p. D. 
- 50 -
Moreover, this restricted view of the role of 
the social sciences can al so be seen in the use of the 
term "social" araong Al researchers. For example, in 
Boden's article (1984) about Al and social forecasting the 
follovrlng stateraents can be found: 
"Put to commercial use, Al-programs will 
appreciably affect not only markets, but also 
personal and communal life-styles. Expert systems, 
for instance, will raise legal, social, and 
psychological problems of an unfamiliar kind. 
Whether they are used to make decisions or merely 
to provide expert advice to (probably less expert) 
decision-makers, the status of the human expert 
will inevitably be affected. And, on the 
international level, their use in countries 
lacking the relevant expertise may be seen 
ambiguously as helpful or exploitative - much as 
human technicians are" (p. 347). 
In such views, "social" has to do with the 
effects of Al, but not with its genesis. 
According to Woolgar (1985), some sociologists 
have also adopted a similar restriction of "social*1 in 
their treatment of Al. They are mainly interested in 
topics such as social attitudes to Al, public perceptions 
and acceptability of machine intelligence, and the likely 
effects of the implementation of Al in different 
institutional environments. At this point, the follovrlng 
question can be raised: can a sociologist vithout a 
detailed consideration of the process of the Al research 
itself (e.g., research into structure of knowledge and 
knowledge representation schemes) discusses the irapact and 
effects of Al? I think not. 
It can also be argued that this reduction of 
sociological capability has no legitimate theoretical 
- 51 -
background because it corresponds to the pre-Kuhnian view 
of science. Concoraitant with that outdated view is a 
distinction between the "technical" (soraetimes 
"intellectual" or "cognitive") aspects of science, on the 
one hand, and peripheral "social factors" on the other. 
This distinction was regarded as definitive of the 
scientific enterprise; "social factors" were precisely 
those factors not related to "science itself"; the domain 
of the "social" was regarded as outside or (at best) 
peripheral to the actual science. But recent work (i.e., 
post-Kuhnian sociology) has established that our 
understanding of science need not be so restricted; the 
nature and content of scientific knowledge is now 
recognized as a legitimate sociological object. Or, to put 
it into the Al context: 
"Sociological studies which focus solely on the 
impact of Al research, to the exclusion of the 
research activity itself, similarly underwrite the 
distinction between the scientific and the social" 
(Moolgar, 1985, p. 560). 
With regard to the above arguments, two-sided 
exclusion of sociology from Al research can be discussed: 
1 - Al researchers have not been interested in 
possible contributions of sociological research into 
crucial Al issues, for example knowledge representation; 
2 - the function of the sociology has been reduced on 
the investigations of impact and effects of Al, but not 
the Al research activity itself. 
With regard to the first notion, it can be said, 
however, that there are more and more articles in the 
- 52 -
literature, which argue strong!y that the reduction of the 
relationship between Al and the social sciences especially 
the cognitive science (i.e., "coraputational metaphor") is 
misleading or mistaken. For instance, Coulter (1985) 
argues that it is confusing and inappropriate to use teras 
and expressions borrowed from coraputer science to explain 
human agencv and social action. 
The next opponent of the "computational 
metaphor" is Bateman (1985). His opposition to the 
cognitive science view, the goals of which are evident 
from Dennetfs (1979) statement, 
"We want to be able to explain the intelligence of 
man, or beast, in teras of his design, and this in 
turn in terms of the natural selection of this 
design" (p. 12), 
can be recognized in the following sentences: 
"The intelligence of man is, indeed, to be 
explained in terms of his design, but that design 
is not first and foremost the design of the 
biological entitv. Human intelligence, perhaps as 
opposed to the intelligence of "beasts", is 
primarv a social phenomenon - not one of the sub-
personal psychology" (Bateman, 1985, p. 78). 
Bateman suggests that some of the central topics 
of Al, such as knovledge representation and planning, that 
have so far been linked most closely with psychology, 
could more fruitfullv draw on sociological investigations. 
Or, in his words, 
"In opposition to the view that knovledge 
representation has anything necessarily to do with 
sociology, I would like to suggest that knowledge 
representation ... is already and necessarily a 
sociological investigation. The main goal of ... 
cognitive science must be to articulate the 
"structures" and "processes" of the human life-
world, not the processes of the hypothesised sub-
personal psychological reality" (p. 65). 
- 53 -
A sirailar approach can also be found in Stamper 
(1985) who argues that the way forward in Al research is 
to treat language and knovdedge as predominantly social 
constructs : 
"Once intelligence has evolved to the level of 
knowledge based on language," its social aspects 
must surely dominate. ... However, in the vrorld of 
Al, computational linguistics and cognitive 
psychology, language arises from some innate 
individual faculty for manipulating syntactic 
structures (p. 172). ... Adopting a simple, 
objectivist view of knowledge does eliminate many 
difficulties from knovdedge engineering, but it 
does lead to people talking of knovdedge as a kind 
of platonic substance which computers can process. 
... Knowledge does not simply exist in a vacuum ; 
someone knows it" (p. 173) 
In this context, it is very important to stress 
that Stamper rejects conventional logics, and proceeds to 
develop an alternative scheme, i.e., logic of action, 
which permits the handling of tirne, space, and context, 
issues vhich traditional logics have found difficult to 
deal vdth. Stamper (1985) emphasizes that, 
"Most mathematicians and logicians seem happily to 
concern themselves with a vrorld of platonic 
reality vvhere no one does anvthing, a vrorld of 
timeless existence. More usefully perhaps, they 
deal vdth the rules for manipulating symbols and 
with the legitimacy of substituting one formula 
for another. But, when it ccmes to relating their 
paper and pencil formalisms to the world of 
practical affairs, mathematicians and logicians 
seem to do no better than their counterparts who 
program computers" (p. 174). 
Stamper's notion lies at the "heart of the 
problem". The fact is that the techniques used for 
representing knovdedge are relatively good at describing 
logical formulas and deductive necessities, and also 
- 54 -
hierarchies of objects. But coverage is particuiariy weak 
for the these other kinds of knowledge: tirne, space, 
events, and action. Therefore, the importance of 
developing new logics of this kind - Stamper's interest in 
this subject grew out of a practical study of how to 
design business inforaation systems; i.e., in the world of 
practical affairs, where one judges the meaningfulness of 
infonnation by its relationship to action - is that they 
may be used as the foundation for Al research that is more 
sensitive to the socially organized and public character 
of human action and cognition. 
Consequently, this leads to a conclusion that 
ali discussions about Al should derive from the analysis 
of the knowledge structure of the applied domains. To 
assess the potential of particular Al applications it is 
necessary to make a distinction between areas where 
knowledge can be represented in a highly structured way 
(e.g., chemistry, mathematics) and fields with "weak" 
formalism (e.g., different domains in social sciences) . It 
is much easier to apply current Al techniques to the 
former domains, for the latter it is necessary to develop 
much more flexible methods. It follows that much more 
fundamental progress should be made in disciplines such as 
psychology and sociology of knouledge to understand the 
whole knovdedge complex (transfer of knowledge, 
formalization of knowledge, common sense, etc). 
This claim, together with the argument that 
knovdedge is a social concept, is already related to the 
- 55 -
second notion; or in other words, the distinctions between 
the "social" and the "scientific" which is a major barrier 
to a thoroughgoing sociological analysis of Al, "need to 
be transcended" (Vfoolgar, 1985, p. 559). 
Therefore, it can be said that Al and sociology 
cannot profit frcra each other, as long as the relationship 
between Al and social sciences is reduced to a 
"computational metaphor", i.e., on the investigations of 
cognitive psychology into human learning and meraory; and, 
as long as the contribution of sociology is seen to lie in 
discussions about the effects and impact of Al. This means 
that when sociology is asked to assess the effects of Al 
it should not only rely on the claims in Al literature, 
but should also investigate the practical day-to-day 
activities of Al researchers, for example, their approach 
to knovdedge representation. I think that this notion also 
explains why there are so many discrepancies in reports 
about the state of the art. 
In the follovdng chapter, the need for a broader 
social sciences investigation into the main issues of Al 
(especially knovdedge representation) will be explained by 
the example of expert systems where it is much easier to 
illustrate the concepts and techniques then in Al in 
general. Questions such as, can the knovdedge of an expert 
be encapsulated in logical scheraes, etc., will be raised, 
and teras such as "taci t" knovdedge will be introduced. 
This will enable the discussion of, on the one hand, the 
- 56 -
potential and liraitations of expert svstems, and, on the 
other hand, the definition of some benefits deriving from 
knouledge representation research. 
- 57 -
Chapter 2 
Expert systems 
"PROSPECTOR has discovered a raolybdenum deposit 
vnose ultimate value will probably exceed 
$100,000,000" (F. Hayes-Roth et al., 1983, p. 6). 
"Unfortunately, this particular statement, which 
is similar to others we have encountered 
elsewhere, has no factual basis. ... PR0SPECT0R»s 
success to date has been scientific rather than 
economic" (R.O. Duda, P.E. Hart, R. Reboh, 1985, 
p. 359-360). 
2.1. Variations in assessments of expert systems 
It has already been said that the development of 
expert systems programs is one of the results of the shift 
in Al research to a knowledge-based approach. Expert 
systems are also known under the name knowledge-based 
systems. The fundamental assumption in expert systems is 
nknowledge is power" (Buchanan and Duda, 1983, p. 165), 
because the specific knowledge of the task (usually vithin 
the narrow area of expertise) is coupled with general 
problem-so Ivi ng knovrledge to provide expert-level analysis 
of difficult situations. 
Unfortunately, it seems that the view on the 
development of expert systems, i.e., the shift from 
problem-independent solution methods to problem specific 
knovrledge, is the only point of agreement between authcrs 
who try to describe and analyse this highly controversial 
- 58 -
topic. In ali other aspects, starting frora the basic 
issues, for exaraple, definition, aims, architectural 
principles, practical use of expert systems, etc., it is 
very difficult to find any consensus. In addition, there 
is also considerable confusion in the terrainology used to 
describe expert systems. This lack of consensus, of 
course, produces many difficulties when trying to define, 
describe, and assess this field. One of the most worrying 
questions is how the readers and potential users of expert 
systems can rely on these controversial statements. The 
greatest danger is in the uncritical adoption of claims 
frcm "popular" literature. To support the arguments about 
the uide variations in the field of expert systems, I 
would like to give some examples. 
1 - disagreements about the definition of expert 
systems; a major problem when studying expert systems is 
the lack of clear definition of what they are. Bramer 
(1981) describes them as computer systems, which embody 
organized knovledge concerning some specific area of human 
expertise, sufficient to perform as skilful and cost-
effective consultants. Sovdzral (1985) sees them as 
computer programs modeled after human experts; they solve 
problems by mimicking human decision-making processes. 
Further, other authors (e.g., Denning, 1986) claim that 
expert systems are, after ali, nothing more than computer 
programs. In contrast, many authors try to avoid 
definition of expert systems with exp!anations about what 
- 59 -
distinguishes an expert systera from an ordinary computer 
application (see Nau, 1983; Yaghmai and Maxin, 1984). I 
think that we could provide an endless number of such 
contradictory definitions. 
Further problems are the preponderance of 
synonyms and the raisunderstandings of basic teras. Expert 
systems are often referred to as knowledge-based systems, 
pattern-directed inference systems, problem-solving 
systems, etc. Two extreme examples are in the equating of 
expert systems, on the one hand, to rule-based systems 
(for criticism see Bramer, 1985), and, on the other hand, 
to intelligent knowledge-based systems (for criticism see 
Alvey Programme, 1985). 
On this ground, one can agree with the 
statement, vrritten by Cendrowska and Braraer (1984) who 
claim, 
"... no ... universally accepted theory exists at 
the present tirne, nor even a universally agreed 
definition of the term Expert System" (p. 229). 
2 - dlsagreements about evaluations of expert systems: 
the fact is that expert systems have been widely acclaimed 
as the applied end of Al research, the long-awaited 
tangible outcome of research investment. Because the 
activity is finding a major commercial marke t (e.g., "One 
could imagine some use for expert systems in just about 
any sphere of business, engineering or research" - Webster 
and Miner, 1982, p. 60), it is often hailed as the 
justification and ultimate application of many years of 
endeavour in Al. 
- 60 -
But the extraordinary optiraism of some reports 
is elsewhere countered by considerable caution and 
pessiraism about the achievements to date. There are some 
niče examples of these disagreeraents: 
a) claims about DENDRAL: "The DENDRAL svstem is nov/ 
in daily use by chemists at Stanford as well as by others 
in universities and industry" (Bramer, 1981), vs. "The 
DENDRAL and META-DENDRAL program are not used outside 
Stanford University and represent rather a demonstration 
of scientific capability" (Belkin and Vickery, 1985); 
b) MYCIN, defined as most significant expert 
system, was never used by doctors for whom it was 
designed; ironically, this essential fact is very rarely 
mentioned in reports about expert systems; 
c) R1, the system used to configure the VAX 
mainframe (DEC reports a $10 millions annual savings - see 
Haves-Roth, 1985a), has only recently come under criticism 
for being muc h more difficult to amend than a 
straightforward program would be (see Leith, 1986); 
d) PROSPECTOR is the subject of many claims about 
savings it had made for exploration companies (see Haves-
Roth et al., 1983); alarmed by these claims, the designers 
of PROSPECTOR wrote to Artificial Intelligence (see Duda, 
Hart, and Reboh, 1985) and pointed out, that;, on the 
contrary, PROSPECTOR had made no savings at ali. . It had 
never even been used as the basis of any exploration plan. 
- 61 -
These are only some of the exampies which 
illustrate wide variations in clairas about the field. 
These disagreements could be analysed on the additional 
topics, for exaraple: 
- what are the essential characteristics of expert 
systeras (performance criteria or architectural 
principles)?; 
- who are the users of expert systems (only experts or 
also naive end-users)?; 
- what are functions of expert systems 
(interpretation, diagnosis, design, etc.)?; 
- what is the aim of building expert systems (to 
duplicate intelligent human behaviour or only to assist 
experts)?, etc. 
Ali these disagreements indicate that the field 
is stili in a state which can be described as "pre-
paradigmatic" in the terminology of Kuhn (1962), with many 
problems left to solve before expert system building, 
"... can emerge as a science rather than the craft 
it is nou" (Bramer, 1985, p. 3). 
At this stage the main task is to find the 
reasons which underpin these disagreements. But before 
undertaking such analysis, it is necessary to develop a 
cocnplex view on expert systems, which will indicate an 
additional feature in expert systems descriptions, i.e., 
over-simplifications of the field. 
The tendency for simplified descriptions derives 
mos tiy frora research in Al, a field in which the canplaint 
- 62 -
is often made that published accounts of research 
frequently do not directly correspond to actual vorking 
prograras and often give a misleading impression of what 
has been achieved. This problem nas already been mentioned 
in the context of the AM program (see Ritchie and Hanna, 
1982). 
Expert systems programs are often large and 
extremely complex, and thus not usually suitable for 
publication. Inevitably, 
"... published accounts tend to be in simplified 
form and this effect is compounded as second- and 
third-hand versions appear in textbooks and survey 
articles" (Cendrovska and Bramer, 1984, p. 230) 
There is no need to stress the relevance of 
textbooks in passing on scientific knowledge to newcomers 
to the subject. To illustrate this notion the MYCIN 
program will be used, which will also serve as an 
introduction in to the field of expert systems. 
2. 2. Oversimplified descriptions of main issues in expert 
systems (a MYCIN čase study) 
MYCIN is a rule-based expert system developed by 
E. Shortliffe (for detailed description of the system see 
Shortliffe, 1976; Cendrovska and Bramer, 1984) at Stanford 
University in the early 1970s. It vas designed to assist 
physicians in the diagnosis and treatment of blood and 
meningitis infections. It tries to model the chain of 
- 63 -
reasoning used by the specialist by embcdying his 
judgmental knowledge in the form of production rules, or 
in other words: 
IF (condition) THEN (iraplication) 
It is claimed by Buchanan and Duda (1983) that, 
although MYCIN is now several years old, it is 
representative of the state of the art of expert systems 
in its external behaviour. 
At this point, the first problematic guestion 
can be stressed: can a program which was never put into 
routine use in hospitals, be defined as representative of 
the state of the art? 
MYCIN collects relevant information about the 
patient (e.g, his clinical conditions, symptoms, medical 
history or details of any laboratory findings) by 
conducting an interactive dialogue with the physician. It 
asks the basic questions first, for example (13) - user's 
responses follow double asteriks: 
PATIENT-248 
1) Patienfs name: (first-last) 
•* Dick H. 
2) Sex: 
** MALE 
3) Age: 
•• 52 
4) Have you been able to obtain positive 
raicrobiological information about a possible 
infection of Dick H.? 
** YES 
INFECTION-1 
5) What is the infection? 
•• ENDARTERITIS 
- 64 -
It is claimed by Daviš (1981) that this style of 
interaction is sirailar to what goes on in hospitals. But 
however good MYCIN's style of interaction and its 
explanation facilities (it can answer to how and why 
questions) , when discussing the possible use of MYCIN in 
hospitals, it is said by Adler (1984), 
"I have not met a doctor yet, ... whose ego is 
small enough to handle that" (p. 262). 
Can the answer to why it is so difficult to put 
medical expert svstems into routine use be found here 
perhaps? Or it is necessary to analyse the structure of 
knowledge in MYCIN and see if ali components of expert 
knowledge are encapsulated in the program? 
The interview continues: 
The first significant organism from this blood 
culture (CULTURE-1) will be referred to as: 
ORGANISM-1 
9) Enter the identity of ORGANISM-1: 
** Unknovm 
10) Is ORGANISM-1 a rod or coccus (etc.): 
«* ROD 
11) The gram stain of ORGANISM-1: 
*» GRAMNEG 
12) Have there been POSITIVE cultures yielding 
organisms about which you will NOT be seeking 
advice? 
«« YES 
Here, it is iraportant to stress that MYCIN 
accepts "UNKNOWN" as a legitimate value. It does not need 
complete, or even correct, data for ali ansvers. There are 
al so many other features of MYCIN which cannot be 
- 65 -
illustrated here, for exaraple: 
the system is tolerant of spelling or typing 
mistakes and can recognize synonyms; 
it communicates with the user in a subset of 
English. 
As it will be seen, the information entered by 
the physician in response to these questions is used by 
the rules in an atterapt to make a diagnosis. After between 
thirty and as many as eighty or ninety questions in one 
consultation, the physician sees a diagnosis similar to 
this: 
INFECTION-1 is ENDARTERITTS with BACTEREMIA 
<Item 1> E. COLI [ORGANISM-1 ] 
<Item 2> SALMONELLA (species unknovm) 
[ORGANISM-1] 
<Item 3> KLEBSIELLA-PNEUMONIAE [ORGANISM-1] 
<Item H> PSEUDOM0NAS-AERUGIN0SA [ORGANISM-1] 
<Item 5> ENTEROBACTER [ORGANISM-1] 
<Item 6> PROTEUS-NON-MIRABILIS [ORGANISM-1] 
If, during this process, further information is 
required, the system will either try to infer it frcm the 
data it has already acquired, or it will ask the physician 
for it. As soon as a reasonable diagnosis can be made, 
MYCIN will compile a list of possible theraples and, on 
the basis of further interaction with the physician, will 
choose the most appropriate one for the patient. MYCIN 
prints out these comments: 
- 66 -
[Rec 1] My preferred therapy recoraraendation is as 
follows: 
In order to cover for Iteras <1 3 4 5 6>: 
Give: GENTAMICIN 
Dose: 128 mg (1 .7 mg/kg) q8h IV (or IM) for 
10 days 
Comments: Modify dose in renal failure 
In order to cover for Item <2>: 
Give: CHLORAMPHENICCL 
Dose: 563 mg (7.5 mg/kg) q6h for 14 days 
Comments: Monitor patienfs white count 
Do you vdsh to see the next choice therapy? 
*« N0 
In this čase MYCIN recommended two medicines to 
treat ali the possibilities. 
When MYCIN is described and analysed in 
textbooks and articles it is often stressed that its most 
important features are as follows: 
- a backward chaining inference system to reason 
"backwards" from diagnosis to symptoms; 
- the use of rules with ncertainty factcrs"; 
- an "explanation" facility to justify the inferences 
made by the system, ccmbined with a means for the expert 
user to refine the system's knowledge base if deficiencies 
are found. 
On the following pages I will illustrate some of 
the characteristics of MYCIN as they are presented in the 
"popular" literature, and at the same tirne, show how these 
descriptions are often over-simplified. i In addition, to 
provide a complex view on expert systems, some other 
issues which cannot be encompassed in MYCIN (e.g., fuzzy 
sets) , will al so be outlined. 
- 67 -
2.2.1. Architectural principles 
The first of the over-siraplified views on MYCIN 
can be found in the claira that its architecture is very 
simple, as illustrated in Fig. 6.: 
Expert svstem 
Inference engine 
«nowledge base 
Fig. 6.: Structure of MYCIN (after Daviš, 1984, p. 33) 
According to such views, the knowledge base 
contains everything that is known about infectious disease 
diagnosis and therapv. The inference engine does the 
computation, taking knowledge frora the knowledge base and 
putting it to work. Many authors al so stress that in the 
čase of expert svstems in general it is important to think 
about the knowledge base and the inference engine 
- 68 -
separately because this should promote flexibility and 
transparency (i.e., the knovledge base can then be 
manipulated like any other data structure). In addition, 
the same inference engine can be kept even when a new 
domain requires a new knovledge. With this goal, as an 
extension of the MYCIN project, a subject-independent 
version known as EMYCIN ("empty MYCIN", or "essential 
MYCIN") has been set up, by removing the detailed medical 
information frcm MYCIN, whilst leaving the overall 
"backvard chaining" framevork, explanatory capabilities, 
etc., intact. With regard to these claims, it is necessary 
to sound a word of caution, and to outline some of the 
serious over-simplifications, starting with the knowledge 
base. 
The knovledge base in expert systems is often 
defined as a body of knowledge specific to the problem 
area (e.g., meningitis) that the system is set up to 
solve; this knovledge is stored in seme manipulatable 
form, by use of suitable formalisra, or knovledge 
re presentation. 
MYCIN's knovledge base, vhich contains about 450 
rules, is often deseribed like this: 
Rule 085 
If 1) the stain of the organism is gramneg, and 
2) the morphology of the organism is rod, and 
3) the patient is compromised host, 
Then There is suggestive evidence (.6) that the 
identity of the organism is pseudomonas . 
- 69 -
The intention of such exarapl.es is to show that 
knowledge is contained in simple if-then inferential rules 
(also known as production rules) , i.e., if we know that 
certain conditions have been met, than we can make certain 
conclusions. But, it is often "forgotten" to add that, 
a) each rule has both an internal (stored) form and 
an external English translation. In the internal form, 
both the premise and the action part of a rule are held as 
a (LISP) list structure. The internal form of the above 
example is then as follows: 
Rule 085 
Premise ($AND (SAME CNTXT STAIN GRAMNEG) 
(SAME CNTXT MORPH ROD) 
(SAME CNTXT C0MPR0M T) 
Action (CONCLUDE CNTXT IDENT PSEUDOMONAS TALLY .6) 
b) although MYCIN's knowledge is largely rule-
based, there is, according to Cendrowska and Bramer (1984) 
an important component, namely the creation of a list of 
potential therapies and the choice of the apparent first 
choice drug, which is algorithmic in nature. 
c) MYCIN's architecture is much more complicated, 
and the entire MYCIN system comprises three subprograms 
(see Fig. 7.): the consultation program, the explanation 
program, and the rule acquisition program. It stores its 
information in two databases: a static database which 
contains ali the rules used during a consultation, and a 
dynamio database which is created afresh for each 
- 70 -
consultation and contains patient information and details 
of the questions asked in the consultation to date. 
Fig. 7« Flow of control and information vithin the MYCIN 
system. Flow of control is indicated by heavy 
arrows, flow of information by light arrows 
(after Cendrowska and Bramer, 1984, p. 233). 
2.2.2. Inference methods 
The inference engine of expert systems raakes 
decisions about hov/ to use the system's knovledge by 
organizing and controlling' the steps it takes to solve 
current problems. Thus, inference methods are closely 
coupled to knowledge representation schemes. According to 
- 71 -
Buchanan and Duda (1983), data-driven control and goal-
driven control are the two main control methods in rule-
based systems. 
With data-driven control rules are applied 
whenever their left-hand side conditions are satisfied. To 
use this strategy, one must begin by entering information 
about the current problem as facts in the database. Here 
it is assumed that a rule is applicable whenever there are 
facts in the database that satisfy the condition in its 
left-hand side. Data-driven control is also knovm by names 
"bottom up", "forward chaining", "pattern directed", etc. 
Its main advantage is in the quick response to input from 
the user. The potential inefficiency of this strategy is 
if that more than one rule is applicable, there is the 
problem of deciding uhich one to apply. 
This problem can be avoided by using a goal-
driven control strategy which focuses its efforts by only 
considering rules that are applicable to some particular 
goal. This strategy has also been used in many systems, 
and is variously known as wtop-down", "backuard-chaining" , 
Mconsequent reasoning", etc. A primary virtue of this 
strategy is that it does not seek information and does not 
apply rules that are unrelated to its overall goal. 
Data-driven and goal-driven strategies represent 
two extreme approaches to control. Various mixtures of 
these approaches have been investigated in an attempt to 
secure their advantages while minimizing their 
disadvantages (14). 
- 72 -
The popular impression of MYCIN is that of a 
systera with a simple backuard-chaining control structure 
acting on a body of rules. For example, if the program was 
trying to deduce the identity of an organism (see the 
above example) , one of the rules invoked vrould be Rule 
085. 
But the first problem with MYCIN control 
strategy is that interactions are too time-consuming, and 
this can be unacceptable when rapid, real-time response is 
required. 
In addition, according to Cendrowska and Bramer 
(1984), there are many derivations frcm simple backward 
chaining, in particular the use of antecedent rules, self-
referencing rules, mapping functions, etc. 
2.2.3. Unreliable data or knov/ledge 
One of the characteristics of an expert's work 
is "reasoning under uncertaintyn. Ebcperts scmetimes make 
judgments under pressure of tirne; ali the data may not be 
available; some of the data may be suspect; some of the 
knowledge for interpreting the data may be unreliable. The 
general problem of dravdng inferences from uncertain or 
incomplete data has led: to a variety of technical 
approaches in expert systems building. 
One of the earliest and simplest approaches to 
reasoning with uncertainty is the use of numbers, called 
certainty factors which indicate the strength of a 
- 73 -
heuristic rule. This approach was also used in the MYCIN 
systera. 
In the čase of MYCIN a certainty factor is a 
nuraber between -1 and +1 and is used to indicate the 
degree of belief that the value of the clinical parameter 
is the true value. A certainty factor of +1 indicates that 
the parameter is Mknown with certainty" to have that 
value. A certainty factor of -1 indicates that the 
parameter is known with certainty not to have that value. 
Certainty factors can be either computed or entered by the 
physician. 
This is also one of the questionable features of 
MYCIN, although it is often claimed in the literature that 
handling uncertain information is one of the main 
advantages in expert systems building. According to Adams 
(1976), there are interdependence restrictions which need 
to be applied to the estimation of certain parameters 
("measure of belief" and "measure of disbelief" in a 
hypothesis, supplied by the physician) to maintain 
internal consistency, but which are not included in the 
MYCIN model. In addition, the use of certainty factors as 
a means of ranking hypotheses is also suspect, since 
examples can be given of cases where, of two hypotheses, 
the one with the lower probability would have the higher 
certainty factor. On the basis of Adam1s analysis, it 
would seem that the MYCIN model has serious limitations. 
Another approach to inexact reasoning, very 
popular in the last few years, is based on the theory of 
- 74 -
fuzzy sets, first proposed by Zadeh (1979). In contrast to 
conventional set theory where the function "belongs to" 
assigns objects to sets, in fuzzy set theory a "degree of 
belonging" may be specified when making this assignment. 
For exaraple, were one to define two sets "tali people" and 
"short people", given a person six feet tali one vrould 
assign him wholly (that is, using a multiplier of one) to 
the tali class. However, where he five feet nine inches 
tali the notion "quite tali" may be expressed by applying 
a belonging multiplier of 0.8 for the tali class and 0.2 
for the short class. This multiplier notion should not be 
confused vdth the idea of a probability. There is no 
probability involved in saying that someone is quite tali. 
This approach is appropriate for areas where subtle 
distinctions are needed between objects. 
The whole topic of reasoning under uncertainty 
is reviewed by Buchanan and Duda (1983). 
2.2.4. EMYCIN as introduction to "shells" 
It has already been said that EMYCIN was 
developed to provide a frameviork in which other systems 
can be built. The potential value of using EMYCIN is well 
illustrated by Feigenbaum (1979). In describing the 
development of PUFF, an expert system for the diagnosis of 
pulmonary function disorders constructed within the EMYCIN 
framework, Feigenbaum points out that the development time 
taken to reach a working system based on the analysis of 
- 75 -
some 250 test cases was less than 10 man-weeks of effort 
by knowledge engineers, with less than 50 hours of 
interaction with subject experts. 
But, apart frctn the fact that using EMYCIN as a 
standard fraraework can greatly reduce the development tirne 
of an expert system, there are weaknesses which may often 
not be apparent to the user. One of the restrictions is 
the overall representation chosen (i.e., contexts, joined 
in a tree structure). Commenting on this representation, 
Van Melle (1980) remarks: 
"EMYCIN vas not designed to be a general pur po se 
representation language. It is thus wholly 
unsuited for some problems. ... The frameviork 
seems well suited for some deductive problems, 
notably some classes of fault diagnosis. ... It is 
less well suited for "formation" problems, ..., 
simulation tasks, and tasks involving planning 
with stepwise refinement" (cf. Bramer, 1985, p. 
4). 
Van Melle goes on to discuss the system's 
backward chaining form of reasoning and canments: 
nBackward chaining works well for simple deductive 
tasks, but as soon as a problem poses additional 
constraints on the means by which reasoning can or 
should be performed, more sophisticated control 
mechanisms may be reauired" (cf. Bramer, 1985, 
p.6). 
There is no doubt that the choice of a suitable 
control mechanism is one of the key problems of Al 
programming. Therefore, the same comments should be made 
for ali commercially-available "shells" which have 
appeared in the last few years and are based on either 
DiYCIN or its fairly close relative PROSPECTOR. 
- 76 -
2.2.5. Explanation facilities 
It is often claimed that explanation facilities 
are one of the greatest advantages of expert systems over 
ordinary coraputer programs (see Dud a and Shortliffe, 
1983). Michie and Johnson (1985) go even one step further 
and speak about the need to develop a human window as the 
opposite to the lunatic black box. 
In the čase of the MYCIN explanation facility, a 
physician can ask why the question the program asks is 
relevant. The response will be something like this: 
14) Is the patient's illness with ORGANISM-1 
a hospital-acquired infection? 
»* Why? 
The user asks "Vfttt" to see an explanation of the 
line of reasoning MYCIN is follovring. The program first 
rephrases the question and then shows the rule whose 
premise it is trying to establish. 
[i.e. WHY is it important to determine whether the 
infection vath ORGANISM-1 vas acquired while the 
patient vas hospitalized?] 
[2.0] This will aid in determining the identity of 
ORGANISM-1. 
It has already been established that 
[2.1] the morphology of ORGANISM-1 is rod 
[2.2] the gram stain of ORGANISM-1 is gramneg 
[2.3] the aerobicity of ORGANISM-1 is facul 
Therefore, if 
[2.4] the infection with ORGANISM-1 was acquired 
while the patient was hospitalized 
then 
there is weakly suggestive evidence ( .2) 
that the identity of ORGANISM-1 is 
pse udom onas [RULE 050] 
MYCIN was using rule 50 and already had certain 
information about the patient. If the physician adds this 
- 77 -
fact, MYCIN will know whether this rule will fire. If it 
does fire, MYCIN will have more information about the 
topic that it is dealing with. 
In addition to interrupting a consultation, the 
physician can wait until the end and ask MYCIN hov it 
reached its conclusion: 
•• HOW DID Y0U DECIDE THAT ORGANISM-1 WAS AN E.COLI/ 
I used rule 084 to conclude that the identity of 
ORGANISM-1 is e.coli. This gave a cumulative c.f. of 
(.51). 
I used rule 003 to conclude that the identity of 
ORGANISM-1 is e.coli. This gave a cumulative c.f. of 
(.43). 
Therefore, by keeping track of the rules that 
vere used, MYCIN can describe hov it reached its 
conclusion. 
As it can be seen, MYCIN is one of expert 
systems with quite dynamic explanation facilities. It is, 
perhaps, useful to stress that MYCIN,s explanation 
capability vas extended by the expert system GUIDON to 
include provision for tutoring. 
Hovever, there are more and more scepticisms 
about the adequacy of current explanation facilities, for 
example: 
"Explanation systems ... usually displayed the 
inference strategy of the system, not the expert. 
We did not explain our reasoning to other people 
in the vay expert systems did. Explanation systems 
do not mirror hov people talked to each other. ... 
Current explanation systems appeared to be good 
for finding out hov the system had come to its 
decision, e.g., for debugging purposes. But as a 
method for explaining information to a naive user 
they vere quite poor" (Alvey Programme, 1985, p. 
117). 
- 78 -
This discussion has se far shown, on the one 
hand that, when discussing expert systeras, it is always 
appropriate to sound a note of scepticism, i.e., it is 
necessary to have in mind that this field is stili ill-
defined. Or, in other vrords, the foundation on which 
expert systems rest has been deseribed by Sheil (1984) as 
being, 
"... a weak technology with few good boundaries 
and ... atheoretic" (cf. Town et al., 1985, 
Section 3.1., p. 10). 
On the other hand, it has also been seen that 
there are four main issues in the design of expert 
systems: 
- knowledge representation; 
- inference methods; 
- methods for reasoning under uncertainty; 
- explanation facilities. 
Deriving from the idea of knov&edge as the 
central issue in Al research, it will be useful to outline 
the main knowledge representation schemes. 
2.3« Representation of knowledge in expert systems 
Before saying anything about knowledge 
representation schemes I would like to emphasize that this 
section will be quite brief because it is very difficult 
not to beceme a "vietim" of over-simplified approach 
discussed in the previous section. 
- 79 -
It has been repeated many tirnes in this study 
that the fundamental observation arising from work in Al 
has been that expertise in a task domain requires 
substantial knowledge about that doraain. The effective 
representation of domain knovledge is therefore generally 
considered to be the keystone to the success of expert 
systeras. 
In expert systems building the following types 
of representation systems have been used: 
- rule-based systems; 
- frame-based systetns (and semantic nets); 
- logic programming systems. 
Such fraraev*Drks are often called representation 
languages because, as with other programming languages, 
their conventions impose a rigid set of restrictions on 
hov one can express and reason about facts in the viorld. 
At this point it is not my intention to discuss if the 
entire knovdedge of an expert in a particular doraain could 
be expressed in such languages, or to stress the 
importance of understanding the structure of such 
knowledge, but to provide a brief description of these 
three ways of representing knowledge(15). 
2.3.1- Rule-based production systems 
Rule-based production systems, developed by 
Newell and Simon (1972) for their models of human 
cognition, are defined as a modular representation scheme 
- 80 -
that is finding increasing popularity in expert svstems. 
It is claimed by Hayes-Roth (1985b) that they constitute 
the best currently available means for codifying the 
problem-solving know-how of human experts. The basic idea 
of these systems is that the database consists of rules, 
called productions, in the form of condition-action pairs: 
IF (condition) THEN (action) 
The left hand side of the rule describes a condition, and 
the right hand side describes the consequence if the 
condition is met. Once a database of such rules has been 
developed, it is possible to apply them systematically in 
a given context, in effect generating and testing 
hypothesis until one that applies is found. 
As has been described, a typical example of a 
rule-based system is the MYCIN program which contains 
about 450 rules. Another twD very well known expert 
systems which are based on production rules are R1 
(configures the VAX/780 ccmputer and contains over 2000 
rules) and PROSPECTOR (aids geologists in evaluating 
mineral sites for potential deposits and contains about 
1600 rules). In the čase of R1, such a rule in a database 
may read (after Kraft, 1984): 
If The current subtask is assigning devices to unibus 
modules 
and there is an unassigned dual port disk drive = 
and the type of controller it requires is known 
and there are two such controllers, neither of 
which has any devices assigned to it 
and the nuraber of devices which these controllers 
can support is known 
Then Assign the disk drive to each controller and note 
that each controller supports one device. 
- 81 -
It is also claimed that production systeras have 
two characteristics: first, existing knowledge can be 
refined, and new knowledge added, for incremental 
increases in system perforraance (derivations frcm this 
principle have also already been stressed). Second , 
systems are able to "explain" their reasoning. 
However, the horaogeneity and siraplicity of 
expression attained in rule-based systems may, according 
to Fikes and Kehler (1985) reduce the ability to express 
other kinds of knowledge; in particular, their expressive 
power is inadequate for defining terms and for describing 
dcmain objects and static relationships among objects. 
These inadequacies can be handled by another knowledge 
representation technique, i.e, by frames. 
2.3.2. Frame-based systems 
Frame-based systems are the most recently 
developed Al knovdedge-representation scheme. Frames are 
data structures in which ali knowledge about a particular 
object or event are stored together. 
Many different variations have been proposed for 
frame-based knovdedge representation, but most of them 
include the idea of having different types of frames for 
different types of objects, with fields or slots in each 
frame to contain the information relevant to that type of 
frame. Thus a frame for a book description will have slots 
for title, author(s), publication date of the book, number 
- 82 -
of pages, etc. To describe a particular book, a copv of 
this book frame would be created, and the slots would be 
filled in with the information about the specific book 
being described. 
One of the most often cited kinds of frames vas 
developed by Schank (1975) in the context of a "theory of 
conceptual dependenoy" . This, among other things, attempts 
to represent most events in terms of a small numbers of 
primitive actions. Each primitive action may be 
represented by a single kind of frame. For instance, 
Schank's theory casts "take" and "give" as two examples of 
the same phenomenon: a transfer of possession. The frame 
for a transfer of possession is: 
name of frame: 
type of frame: transfer of possession 
source: 
destination: 
agent: 
object: 
where the source is the person or thing from which the 
object is taken, the destination is the person or thing to 
which the object is given, and the agent is the one who 
performs the transfer. When the above frame is ready-made 
for the sentence "Bili took the book from Margaret", the 
result is: 
name of frame: T1 
type of frame: transfer of possession 
source: Mary 
destination: Bili 
agent: Bili 
object: book 
The main advantage of frames is that ali the 
relevant information is collected together, accessing and 
- 83 -
manipulating information is then easier. Several computer 
languages have been or are being developed to provide ways 
to manipulate frames, one of the most popular being KRL 
(see Bobrow and Winograd, 1977). 
Frame representations are basically equivalent 
to semantic nets which were invented as an explicitly 
psychological model of human associative meraory. Semantic 
nets are like frames in the sense that knovledge is 
organized around the object being described, but here the 
objects are represented by nodes in a graph and the 
relations among them are represented by labeled arcs. 
According to Mylopoulos and Levesque (1983) there are 
three advantages of semantic nets. Due to their nature, 
they directly address issues of information retrieval 
(this notion will be outlined in the last chapter). 
Another important feature is the availability of 
organi zational principles. A third is the graphical 
notation that can be used for netvrork knovdedge bases and 
that enhances their comprehensibility. A major drawback of 
netvrork schemes is the lack of formal semantics and 
standard terminology. 
2.3.3. Logic programming systems 
The key idea underlying logic programming is 
programming by description. In traditional softvvare 
engineering, one builds a program by specifying the 
- 84 -
operations to be performed in solving a problem, that is, 
by saying how the problem is to be solved. In logic 
programming, one constructs a program by describing its 
application area, that is, by saying, what is true. A 
description of this sort becomes a program when it is 
combined with an application-independent inference 
procedure. Applying such a procedure to a description of 
an application area makes it possible for a machine to 
draw conclusions about the application area and to answer 
questions even though these ansuers are not explicitly 
recorded in the description. 
One such language is PROLOG (see KowalsIcLf 1977; 
and Clocksin and Mellish, 1981) which works with objects 
and their relationships, specified by the programmer as 
rules. Relationships might be (16): 
John likes Mary 
Philip father-of Charles 
Charles father-of William 
Mary likes John 
New relationships can be defined: 
x friends-with y if x likes y and 
y likes x 
and questions can be asked, such as, is John friends with 
Mary? 
In PROLOG: Does (John friends-with Mary) 
Ansver: YES 
The advantages of PROLOG are evident from the 
fact that PROLOG has been chosen as a standard language 
for the Japanese Fifth Generation Computer Systeras 
Project. In addition, it is al so often claimed that the 
- 85 -
logical scherae is popular because of its very general 
expressive power and well defined semantics. 
However, according to Genesereth and Ginsberg 
(1985) there are the following 1imitations to most logic 
programraing systems: 
- language constructs are very fine grained and do not 
provide adequate facilities for defining more complex 
constructs; 
- generality of the predicate calculus is a barrier to 
the developraent of effective deduction facilities for 
using knouledge expressed in it. 
On the basis of this short description of three 
main ways of representing knouledge it can be concluded 
that no single representation formalisra seems adequate, 
and that each technique has its own strengths and 
weaknesses. To solve this problem, some hybridizations 
have also been developed. 
This review has also shown that knowledge 
representation is a central issue in expert systems. This 
will be helpful in a discussion about the fundamental 
problems, relevant for the social sciences. To complete 
the picture, it is also necessary to say something about 
the developraent of expert systems, and, in this context to 
stress the differences between "classic" expert systems 
and sc-called "shells" as additional example of 
disagreements between authors. 
- 86 -
2. H. Some features of expert systems developrcent: from 
classic expert systems to "shells" 
The earliest developraent in expert systeras began 
in the area of structural chemistry within the DENDRAL 
project (see Lindsay et al., 1980). Organic cheraistry is 
one of the most appropriate fields for expert systems 
development because it, 
"... has a strong fonnalisra, that of the graphical 
structure diagram or its netvrork equivalent within 
the machine, to represent molecular structures, 
and transformations among them" (Town et al., 
1985, Section 3.1., p.1). 
The whole DENDRAL project includes three 
programs: DENDRAL, CON GEN, and META-DENDRAL. 
The major program HEURISTIC DENDRAL was the 
first and is probably the best known expert system (e.g., 
it has recently been described as the grandfather of 
expert systems - see Aleksander, 1984). The program is 
designed for use by organic chemists to infer the 
molecular structure of complex organic compounds from 
their chemical formulas and mass spectrograms (mass 
spectrograms are essentially bar plots of fragment masses 
against the relative frequency of fragments at each mass) . 
The program was developed by E. Feigenbaum, B. Buchanan 
and others in 1965 at Stanford University. 
The program makes use of rules which relate 
physical features of the spectrogram (high peaks, absence 
of peaks, etc.) to the need for particular substructures 
to be present in or absent from the unknown chemical 
- 87 -
structure. Using these constrair.ts, CONGEN (see Carhart, 
1979) produces a list of ali acceptable candidate 
structures. For each of these a spectrogram is then 
coraputed, and a raatching algorithra ranks the candidates in 
order of the best fit between their spectrograms and the 
spectrogram of the unknown compound. While CONGEN resulted 
from the slow speed of DENDRAL, META-DENDRAL (see Buchanan 
and Feigenbaum, 1978) is an atterapt in the automatic 
acquisition of knowledge. 
Another pioneer system which has already been 
described in detail, is MYCIN. In this context, it is, 
perhaps, useful to say that its framework, EMYCIN, has led 
directly to SACON, an advisory program for structural 
analysis in engineering, and to PUFF which analyses 
results of pulmonary function tests for evidence of 
possible pulmonary function disorder. An offshoot of MYCIN 
is TERESIAS (see Daviš et ali, 1977) which concentrates on 
knovdedge acquisition, i.e., it assists in the 
construction of large knovrledge bases by helping transfer 
expertise from the human expert by means of a dialogue. 
MYCIN's simple knowledge structure has 
stimulated other developments directly, where, among 
others, PROSPECTOR and XCON (earlier named R1) deserve a 
brief description. 
- 88 -
PROSPECTOR is a systera developed by R. Duda, P. 
Hart and others at SRI International in California. It is 
intended to aid geologists in assessing the favourability 
of a given region as a site for exploration for ore 
deposits of various types. The user provides the program 
with a list of rocks and minerals observed "in the field" 
and other Information expressed in a rudimentary form of 
English. The program then conduots a "dialogue" with the 
user, requesting additional information where needed. At 
any point, the program is able to provide the user with an 
explanation of the intent of any question. The eventual 
output frcra the program is an indication of the "level of 
certainty" to which the available evidence supports the 
presence of a particular form of deposit in a given site. 
The svstem has a number of different knowledge 
bases, corresponding to different classes of ore deposits. 
The program's knovledge for a particular ore deposit is 
held in the fora of an "inference network" of relations 
betveen field evidence and geological hvpotheses. The user 
expresses his certainty about a piece of evidence on a 
scale -5 to 5, where 5 denotes that the evidence is 
definitely present, -5 that it is definitely absent, and 
zero indicates no information. These are converted 
automatically into probability-like values. However, as it 
has already been said, PROSPECTOR provides a niče example 
of controversial claims about the achievements of expert 
systems. 
- 89 -
XCON (earlier named R1), developed in the late 
1970s, is a rule-based expert systera that configures VAX 
computers (see McDermott, 1982). It takes the 
specifications for a new coraputer installation, deterraines 
the physical layout and interconnection of the coraputer's 
components, checks the resulting configuration for order 
and consistency, and, if necessary, either upgrades 
hardware specifications (introduces a heftier power 
supply) or adds missing components (a cable) . According to 
Duda and Shortliffe (1983) it is now used by the Digital 
Equipment Corporation (DEC) to configure every VAX that is 
sold. It is claimed that it results in a $10 million 
annual saving; and 85$ of configurations are reported to 
be faultless (see Hayes-Roth et al., 1983) • However, there 
are two questionable issues: 
- the system is much more difficult to amend than a 
straightforvrard program would be; 
- vihen tested by the user community, the performance 
of the system declined to the 60$ level (see Daviš, 1984). 
The interest in building expert systeras has been 
widening significantly. The follovang are some of the 
well-known examples of such expert systems: INTERNIST 
(used for diagnosis in internal medicine), SECS (proposes 
schemes for synthesising stated organic compounds), 
DIPMETER (advises on oil-well driling), MOLGEN (assists in 
the design of experiments in molecular genetics) , etc. 
- 90 -
Of course, this list could be much more 
coraprehensive; the main difficulty when producing such 
lists is in the lack of firm evaluation criteria, the most 
worrying being the lack of users' appreciations of 
existing expert systeras. 
The systems described above can also be defined 
as "classic" expert systems, recently being estimated as 
"distrusted" and "disused" (see Pogson, 1986). Such 
statements mainly criticize the building of expert systems 
from scratch which often takes at least five man-years of 
effort (see Daviš, 1984). An alternative is to make use of 
a standard framework or "shell" which enables working 
systems to be developed rapidly. As a result, a number of 
cacmercially available "shells" have appeared in the last 
few years. 
Hovever, expert systems "shells" are again a 
topic of many disagreements. One the one hand, authors 
like Gooding (1986) claim that this is, 
"... the only route by which expert systems can 
make a contribution to mainstream computing" (p. 
39). 
On the other hand, D»Agapeyeff (1 984) found in 
surveving the commercial applications in the UK that it 
was necessary to introduce a new term "simpler expert 
systems" to adequately categorise much activity. His 
conclusion about "shells" and their related commercially 
available expert systems is that, 
- 91 -
"It is necessary to correct the impression, much 
heralded hitherto, that Expert Systems are 
inherently coraplex, risky and deraanding. This 
impression is a handicap to competetive 
developments in the supply and usage of Advanced 
Information Technology" (p. 3). 
There is no doubt that using a "shell" can 
greatly reduce the development tirne of an expert system 
because ali "shells" provide a following basic framework: 
- a means of encoding the domain knovledge; 
- inferencing meehanisms (typically backward chaining) 
for making use of the encoded knowledge. 
Therefore, with the ccmplex programming tasks 
being done by the "shell", the task of building an expert 
systera is greatly simplified and the builder is free to 
concentrate on the knowledge acquisition process. This is 
a much quicker and cheaper process than building expert 
systems from scratch. 
As we have already seen, the first and probably 
the best known "shell" is EMYCIN; the creation of PUFF and 
SACON has al so been mentioned. According to Johnston 
(1986), dozens of "shells" vrorkLng on this general 
principle are now on the market and are achieving some 
success. 
In addition, there is a trend of developing 
expert system "shells" for microcomputers (see survey by 
Guilfoyle, 1986a) . Amongst these the most popular are Xi 
(ušes "if ... then" rules and forward and backward 
chaining) and Guru (integrates expert systems building 
tool with spreadsheet, database raanager, text processor, 
etc.) . 
- 92 -
Despite the commerciai success of expert systera 
"shells", it is necessary to say a word of caution: 
- the problem of an overall knowledge representation 
scheme and "backvrard chaining" has already been discussed 
in the context of EMYCIN - when applying "shells" it is 
always necessary to have in mind the differences betueen 
subject domains; 
- expert systems "shells" are primarily commercially 
oriented, and their builders are not interested in some of 
the "academic" issues of Al, for example, the ability to 
learn, etc, which originally excited researchers in Al. 
Therefore, there is a danger of the transformation of 
expert systems into a "... flabby marketing phrase" 
(Gooding, 1986, p. 39). Or, in other words, 
nAI is a chaos. It's hard to get good researchers 
to work on fundamental problems because the 
ccmpanies are snapping them ali up. Theory has 
stagnated for a moment, and we've lost our 
momentum" (Waldrop, 1984, p. 804). 
At the end of this section it can be concluded 
that this discussion has so far revealed some important 
characteristics of expert systems research. One of the 
most alarming is the lack of agreement araong authors about 
even the basic issues, such as definition, aims, 
evaluation, commerciai use, etc, of expert systems. The 
additional problem is in over-simplified descriptions of 
the field and in its atheoretical foundations. 
At this point the question can be asked : what is 
impeding greater clearness in expert systems discussions? 
- 93 -
Is the reason only in the fact that this work is 
relatively new? I think not. 
In this study, the central role of knowledge in 
building expert systeras has often been repeated. 
Therefore, the answer to the above question can only be 
provided in a detailed analysis of crucial expert systeras 
issues (which can also be defined as fundamental 
problems), such as knowledge acquisition and 
representation, and the connected explanation facility. 
According to their nature, there is no doubt that these 
issues should also be addressed to the social sciences. 
2.5. Fundamental problems in expert systems research, 
relevant to the social sciences 
2.5.1. Knovrledge acguisition 
Knowledge acquisition for expert systems is a 
difficult and time-consuming process. Barr and Feigenbaum 
(1981) describe it as the biggest bottleneck in the 
production of these systems. 
Knowledge acquisition is defined as, 
"... the transfer and transformation of problem-
solving expertise frcra some knovledge source to a 
program. Potential sources of knowledge include 
human experts, textbooks, data bases, and one1 s 
own experience" (Buchanan et al., 1983, p. 128). 
It is claimed by Hayes-Roth et al. (1983) that 
this transformation is the heart of the expert system 
- 94 -
development process. In this context, the role of the 
knowledge engineer is introduced whose function is to 
extract the knowledge from the relevant expert and to code 
this knouledge for input to the coraputer. Through an 
extended series of interacticns, the knowledge engineering 
team (the knouledge engineer and the expert) defines the 
problem to be attacked, discovers the basic concepts 
involved, and develops rules that express the 
relationships existing between concepts. Thus, there are 
the follouing major stages in the evolution of an expert 
system, as identified by Hayes-Roth et al. (1983) and 
shown in Table 3.: 
Identification: 
Conceptualization: 
Foraalization: 
Implementation: 
Testing: 
Determining problem characteristics 
Finding concepts to represent 
knouledge 
Designing structures to organize 
knouledge 
Formulating rules that embody 
knouledge 
Validating rules that embody 
knouledge 
Table 3« Stages in the evolution of an expert systems 
(after Hayes-Roth et al., 1983, p. 24). 
Houever, the lack of emphasis placed on the 
techniques (or problems) of extracting expert knouledge 
and converting it into a suitable form (generally rules) 
in the "popular" literature on expert systems might lead 
to a conclusion that it presents no difficulties. 
Unfortunately, very little is knoun at present 
about hou to extract expertise from an expert and almost 
- 95 -
nothing is on offer as a technique. In other words, as 
discussed in a report by Welbank (1983) who exarained a 
variety of knowledge acquisition techniques for expert 
systems and concluded that this field is, 
"... at a very early stage of development, where 
different experiences are st-ill being gathered, 
and general principles have not emerged" (cf. 
Braraer, 1985, p. 7). 
The problem of knowledge acquisition is 
characterized on the one hand by an expert, unfarailiar 
with expert systems and unable to articulate what 
knowledge he has and how he ušes it to solve problems; and 
on the other hand by a knowledge engineer who may well be 
totally ignorant about the domain of expertise. 
In addition, the most effective methods of 
acquiring knowledge frcm experts, such as observation "in 
the field" or in-depth intervievring, are inherently slow, 
a major problem given that experts' tirne is often in short 
supply. While an experienced team can put together a small 
prototype system in 1 or 2 man-months, the effort required 
to produce a system that is ready for serious evaluation 
is more often measured in raan-years. There is no doubt 
that these methods are often expensive and also prone to 
error. 
Some have argued that the best way to overcome 
the problems associated with traditional techniques of 
knowledge acquisition is to move towards automatic methods 
of rule generation based on an analysis of exaraple cases. 
The essential idea behind rule-induction - based on 
- 96 -
Quinlan's ID3 algorithra (see Ouinlan, 1979) - is that 
given a database of exaraples, machine induction can 
quickly generate a rule base which corapletely accounts for 
ali the examples, and in general this can be performed in 
many different ways. The use of this approach is nicely 
described by Michie (1984) who claims that expert systems 
can now al so be used, 
"... to put the knowledge back into human hands in 
improved form" (p. 342). 
One of the successful examples of using this 
method can be found in a study by Mozetič, Bratko, and 
Lavrac (1983). Using the logic programming language 
PROLOG, the authors collaborated with senior clinical 
cardiologists at the Ljubljana University Medical School. 
The Yugoslav group constructed a computer model of a 
complete and ultra-reliable diagnostic scheme for multiple 
arrhvthmias and their relation to the ECG wave form. The 
system produced new knowledge, although sraall in extent, 
but sufficient to have a use in teaching and as a 
reference text for the specialist. 
However, it has al so been clairaed that automatic 
rule induction has some weaknesses (see Bramer, 1985), for 
example: 
automatic induction may result in a set of rules 
that is formally correct (in the sense of accounting for 
ali the examples given) but which has low predictive power 
for the cases outside the example set; 
- 97 -
the possibility of "noise" in data values and the 
possibility that some necessary attributes (perhaps those 
which are only significant for a fairly sraall number of 
cases) are raissing. 
Such reservations should not reduce the role of 
rule induction methods which are at present one of the 
essential steps in the development of expert systems. 
Another, more general, problem connected with knowledge 
acquisition techniques is the question of whether an 
expert's knov/ledge can be represented in its entirety in a 
computer program. 
2.5.2. Knowledge representation 
It is claimed in the literature that specialized 
knov/ledge encapsulated in expert systems is of two types: 
"The first type is the facts of the dornain - the 
widely shared knov/ledge ... that is written in 
textbooks and journals of the field, or that forms 
the basis of a professor's lectures in a 
classroom. Equally important to the practice of 
the field is the second type of knowledge called 
heuristic knowledget which is the knowledge of 
good practice and good judgment in a field. It is 
experiental knov/ledge, the "art of good guessing", 
that a human expert aquires over years of work" 
(Feigenbaum and McCorduck, 1984, p. 76-77). 
In addition, knov/ledge engineers are supposed to 
know how to ex.tract relevant knov/ledge from an expert and 
hov/ to encode that knowledge in a form amenable to 
mechanical manipulation. 
As it is evident from the above quotation, it is 
thought that the essential idea behind expert systeras 
- 98 -
building is that ali human knowledge can be fully 
exhausted by facts and heuristics. 
A word of caution about such claims can be found 
by Aleksander (1984) who emphasizes that, 
"... there are .... raany more areas ... where 
knowledge cannot be encompassed in simple logical 
or probablistic rules" (p. 134). 
An alternative and very useful view on knov/ledge 
elicitation is explained by Collins, Green and Draper 
(1985) who introduce the importance of so-called 
"cultural, tacit, and skilfull aspects of knov/ledge". 
Their main hypothesis is that when domain expert* s 
knovrledge is elicited and encoded, these aspects of 
knov/ledge are lost. This is explained by a very 
illustrative metaphor: knowledge is like chicken soup with 
dumplings, and the expert system is like a colander; with 
ali knov/n expert systems, the dupralings get transferred 
but the soup is lost. The dumplings are the readily 
explicable facets of knov/ledge such as factual information 
and articulateable heuristics, whereas the soup is the 
context/meaning of the facts and the non-articulated but 
"taken for granted" practices and "ways of going on" in 
practical and theoretical settings, or in other words 
"tacit knov/ledge"(1 7) • This has implications for the use 
of expert systems, i.e., expert systems must rely on the 
users' abilities and their "tacit" knov/ledge to interpret 
the system's advice. Consequently, it is clear that the 
more expert the end-user, the easier it v/ill be to build a 
system that v/ill be useful. 
- 99 -
These authors conclude that the crucial division 
in knouledge is not the separation between facts and 
heuristics, as much work on knowledge elicitation has 
stressed, but between the articulateable and the tacit. 
The promise and developraent of expert systems can be much 
better understood once it is realized that liraits are set 
by the fact that substantial coraponents of knowledge are 
not articulateable. 
But, I believe it is necessary to be much more 
precise when maklng this division in knowledge. Additional 
distinctions should be made between domains where 
knowledge can be represented in a highly structured, 
formalized way (e.g., different domains in chemistry, 
raathematics, etc, where we can find immediate contact 
with common sense, or so-called "tacit" knowledge) and the 
areas where knowledge cannot be represented in strong 
formalism (e.g., different areas in the social sciences) . 
Expert systems can be, of course, much more reliable in 
the former area. It is surprising how infrequently this 
problem is discussed in the literature on expert systems. 
Much of the descriptions and analysis are based on the 
assumption that defining "narrow doraain" is the ultimate 
condition for building expert systems, without taking into 
account the characteristics of the knowledge structure of 
this domain. Starting from the structure of knowledge, 
many confusions and disagreeraents about expert systems can 
be avoided. 
- 100 -
This leads to a conclusion that developraents in 
expert systems depend not only upon advances in knowledge 
engineering, but also on research in the wider fields of 
Al and the social sciences which underpin the complex 
approach to knov/ledge, i .e., transfer of knowledge, 
formalization of knowledge, common sense, etc. 
2.5.3. Explanation facilities 
It has been stressed in this study that the 
explanation facility of most expert systems consists of 
nothing more than printing out a trace of the rules being 
used. These facilities are valuable, not least in the 
possibility that the end user can learn about the 
knowledge domain by interacting with the explanation 
facility (e.g., GUIDON). 
However, many expert system explanation 
facilities cannot fulfil a much more important role. 
Michie and Johnston (1985) put the matter this way: 
"Any socially responsible design for a machine 
must make sure that its decisions are not only 
scrutable but refutable. That way the tyranny of 
machines can be avoided" (p. 69). 
In this context, Michie and Johnston talk about 
a "human window" (as opposite to a lunatic black box) in 
coraputer programming - a window of reasoning that is like 
human reasoning in depth and coraplexity. 
At this point, the question can be raised as to 
whether the tvpe of explanations described above can make 
- 101 -
a machine's decisions refutable. The basis of these 
explanations are rules and encapsulated knowledge. 
Therefore, when discussing the reliability of explanation 
facilities, it is very important to start again from the 
understanding of structure of knovledge and its 
formalization in a particular applied doraain. There is no 
doubt that in doraains where knowledge cannot be 
represented in a highly formalized way, these explanations 
are only condensed fragraents of the expert's knowledge. 
There is a whole host of "tacit" knovledge (see Collins, 
1986) which expert systems cannot handle but vhich is 
essential for the provision of a good explanation: that 
is, an explanation that can be refuted. 
Refutability is important because explanations 
are not simply extras which are provided by expert 
systems. D. Michie has long been concerned with the 
refutability of computer programs both from the point of 
view of producing a good system, and also because non-
refutable systems can cause catastrophes when used in such 
areas as air traffic control, air defence or nuclear power 
(see Michie and Johnston, 1985). 
The doubts about vhether current expert systems 
can provide refutable explanation facilities are expressed 
by Leith (1986), who claims: 
"Unfortunately, it is beginning to seem as though 
expert systems have not been designed in a 
socially responsible way, for they cannot really 
explain the basis for their reasoning in as full a 
manner as the non-expert needs" (p. 15). 
- 102 -
Collins (1986) goes even ene step further in his 
eriticism and argues that current expert systera 
explanation facilities can do nothing but add to the 
friendliness, persuasiveness and seeming authority of the 
formalized knowledge eneoded within an expert system. He 
proposes that an explanation facility which cannot be 
refuted should be banned, 
"... and this means cutting out the explanation 
facility except where the expert system is to be 
used by an expert!" (p. 9). 
I think that it is important to repeat here how 
little has been said in these discussions about the 
strueture of knowledge in a particular domain. 
The fact is that the expert systems ccmmunity 
is, nevertheless, aware of the inadequacy of current 
expert systeras explanations (see Guilfoyle, 1986b). The 
following are some attempts to improve the explanation 
facility which indicate the importance of this coraponent 
of expert systems: 
- use of additional knowledge beyond the system's 
performance rules; 
- user modelling: building up a pieture of the user, 
which can help to tailor output, interfaces, help levels, 
and so on, to the particular user; this is an assessment 
of what the user does and does not know and what he is 
trying to accomplish; 
- inereased use of diagrams which can sometimes offer 
better explanations than text, etc. 
- 103 -
At the end of this chapter it can be conciuded 
that there is no doubt that three fundamental issues in 
expert systeras (i.e., knowledge acquisition, knowledge 
representation, and explanation facility) are also 
relevant concepts in the social sciences and are, as such, 
legitimate subjects of social sciences research. This idea 
is very iraportant: On the one hand, expert systems will be 
able to meet the challenge of general competence and 
reliability if more fundamental progress is made by Al, 
psychology, sociology of knowledge, etc, in understanding 
the structure of knovdedge and the whole knowledge complex 
(e.g., transfer of knowledge, formalization of knovdedge, 
process of reasoning, common sense, etc). On the other 
hand, this understanding can be a starting-point for 
seeing vvhere the development of expert systems is going 
and how it will get there. Only such an established 
framework will enable the social sciences to discuss 
realistically the problems of the impact and effects of 
expert svstems and Al. 
The role and potential of Al and expert systems 
for the librarv/information community can also be 
understood frora this point of view. This question was 
raised in the introductory section and vdll be explained 
in: the follovdng chapter. 
- 104 -
Chapter 3 
The relevance of Al and expert systems research for 
library/information systems 
"The goal of expert systems research is to provide 
tools that exploit new ways to encode and use 
knowledge to solve problems, not to duplicate 
intelligent human behaviour in ali its aspects" 
(Duda and Shortliffe, 1983, p. 266). 
Throughout this discussion it has been 
emphasized that knowledge representation is one of the 
fundamental problems in expert systems building. However, 
without regard to the difficulties in trying to encode 
expert's knowledge in representation schemes, it should be 
stressed that research on knowledge representation methods 
and techniques is one of the vital issues in the whole 
field of Al. The results of such research can be important 
to a variety of scientific and economic endeavours, 
including the design of improved library/information 
systems. 
At present there are two main identified areas 
where the achievements of Al and expert systems research 
can be useful for libraries and information services, as 
follows: 
- expert systems building and knowledge representation 
schemes force a rethink of the methods of organizing and 
representing information and knowledge in databases in 
order to make it dynaraic and interactive; 
- 105 -
- the developraent of expert interraediary systeras as 
front ends to bibliographic databases. 
3. 1. A need for new methods of organizing and 
representing information and knowledge in databases 
It has already been said that information 
science which grew out of the library science with the 
introduction of computers and which is, 
"... concerned with formalizing the process of 
knowledge formulation, organization, codification, 
retrieval, dissemination, and acquisitionM 
(Walker, 1981, p. 348), 
and artificial intelligence are relatively new areas of 
research, each having assumed an independent identity 
within the past thirty years. 
The relevance of much of the research in Al to 
library/in format ion systems seems to be in the middle step 
in the "information transfer cycle", i.e., computerized 
information storage and retrieval. 
But first, to provide a theoretical framevrork 
for a discussion it is necessary to clarify the term 
"database systems". 
According to Town et al. (1985) a database 
system may be considered as consisting of three major 
components. Firstly, there are the records that form the 
body of the database, each record comprising one or more 
data elements. These elements may be of several types, for 
example: 
- 106 -
- numeric data; 
- structural data; 
- textual data. 
Secondly, there are the search mechanisms that 
allow a user to query the database so as to retrieve 
records frcra it. Thirdly, there is the interface to the 
database by means of which a user may specify his or her 
query. 
Therefore, a distinction can be made between 
database systems which provide the user with direct access 
to source data, either a variety of different kinds of 
numerical values (also numerical databanks), structural 
data (e.g., structure-based systems in chemistry), or even 
the complete texts of documents (in the legal area, e.g., 
LEXIS) and, in contrast, bibliographic databases, which 
help the user to identify primary or source documents that 
might have information relevant to his needs and 
interests. 
Although it is important to develop so-called 
"system thinking", as defined by Kornhauser (1983), i.e., 
"... an organized way of linking bits of 
information into networks, trees, modular systems, 
showing the interrelationship between data" (p. 
385), 
in ali three kinds of databases, I vrould like to 
concentrate, in this section, on the possible usefulness 
of knovledge representation research to rethinking the 
organization of information in bibliographic databases. 
The problems of interfaces to database systems will be 
outiined in the next section. 
- 107 -
One of the main disadvantages of bibliographic 
systems is that they only provide pointers to literature; 
the user can make a prelirainary assessraent of the utility 
of a reference frora the title and abstract, but he stili 
has to find the documents and evaluate their contents 
before he can derive Information frora them. This is not a 
straightforward matter like looking up an item in a table 
in a databank. 
The use of bibliographic databases also relates 
to the procedures for classifying or indexing the 
document, i .e. the inclusion of a docunent in a database 
requires that a judgment be made about its content. But, 
however appropriate index terms or thesaurus entries to a 
document (either manual or autcmatic indexing) are, there 
is one main drawback: index terms are words, and they can 
have in isolation many meanings. Consequently, document 
searching often produces much irrelevant material. 
The fact is that, if one wanted to create a 
model with respect to user requests, this model, 
"... needs to be more than just a list of index 
terms, but to be terms in relationships" (Addis, 
1982, p. 302). 
In this context, the most acceptable structures 
are those that best maintain the semantic feature of 
information. Currently, this requirement is best fulfilled 
by the previously described knouledge representation 
scheme, called semantic nets. 
- 108 -
This is one of the pcssibic methods which can be 
used as a complement to the subject oriented approach of 
most abstracting and indexing services. Hjerppe (1983) 
identifies within this notion three different 
applications, i.e.: 
- condensing existing knowledge in stages; 
- organizing existing knowledge to show structure and 
relations of documents; 
- organizing existing knowledge to exhibit lacunae and 
unnoticed links. 
An example of an attempt at the first 
application mentioned is the Hepatitis Knowledge Base 
(HKB) which was created in the USA for medical researchers 
and doctors interested in hepatitis (see Bernstein, Siegel 
and Goldstein, 1980). This database which can be searched 
on-line, comprises information initially extracted from 
forty review articles and then expressed in a series of 
hierarchical statements to fora a consensus of ali the 
available knov&edge about the disease. The problem of 
knowledge acquisition and updating for such a database is 
described by Walker (1981). 
Examples of the second and third applications 
mentioned can be found in the co-citation clustering 
concept. According to Small (1986), a co-citation cluster 
is a bibliometrically defined netvork structure, and the 
hypothesis is that it defines a knowledge structure as 
well as an "invisible college" or social structure. In 
essence, the method pieces together selected passages from 
- 109 -
a variety of sources to form, as far as possible, a 
coherent whole. The sources of text for the narrative are 
the papers which cite the core documents in the cluster, 
and specifically the eontext of citation for those 
docuraents. Using this method it has also been shown how 
paragraphs which cite multiple core documents in a cluster 
can be used to provide an interpretation of the structure 
of the co-citation network map (see Small, 1984). Earlier 
works on the database ISI/BIOMED and recent work in the 
field of education (see Ward and Reed, 1983) indicates the 
need for new approaches in the organization and 
representation of inforaation and knowledge in databases. 
In addition, there are also a nuraber of 
, interesting attempts tovards building expert systems for 
"traditional" library work, such as cataloguing (see 
Davies and James, 1984; Hjerppe, Olander, and Marklund, 
1985), reference services (see Bivins and Palmer, 1981 ) 
and some other important library/infonnation tasks. 
3.2. Expert intermediary systems as front ends to 
bibliographic databases 
It has been eraphasized in the previous section 
that the interface to databases is also one of the 
essential elements in the whole database system, 
particularly because, 
- 110 -
"The flexibility and ease of use of this interface 
will play a large part in detennining the degree 
to which the database is utilised by the intended 
user community: systems that require extensive 
effort to express a query, or which provide little 
feedback during the retrieval operations, are 
unlikely to be used at ali heavily" (Town et al., 
1985, Section 2. 1. , p. 1). 
Therefore, a flexible and user-friendly 
interface is needed. There are two approaches which derive 
frora Al research and could be appropriate to such a type 
of interface: 
- question answering systems (e.g., LUNAR, LADDER, 
e te .); 
expert systems interfaces, i.e., their explanation 
facilities (e.g., MYCIN). 
Both approaches have already been deseribed in 
this study and it has been stressed that their main 
advantage is that they allow communication with a database 
via the user's natural language. Their raain drav/back is 
domain dependence. Another important type of interface, 
not discussed here, is the use of graphical interfaces 
(e.g., applying graphics techniques to numeric and textual 
files; see Michard, 1982). 
When desribing interfaces it is also necessary 
to stress the difficulties connected with the problems of 
access to on-line bibliographic databases. According to 
Town et al. (1985), there are currently over 500 online 
bibliographic databases containing more than 100 million 
citations, while a full text file raay contain several tens 
of billions of characters. But, at the present tirne many 
- 111 -
users are unable to make full use of available computer 
searchable databases as these require knowledge of both 
the mechanisms of performing a search and of the way a 
controlled vocabulary may be used to express a document 
search request. 
Because of these difficulties, online searches 
are usually carried out by professional intermediaries 
instead of end-users. Reference librarians and 
intermediares are needed to help formulate user requests 
in terras of the information systems and to provide 
guidance on how the systera is organized, on what materials 
are available, and on how to search for and locate the 
desired items. This is a very questionable situation 
because it is very difficult to determine what the user 
really requires and the searcher may seldcra be aware of 
his own real needs. The connection is, in Pollitfs words 
(1986), that coraputerized searching services will not have 
their full irapact upon user communities until direct user 
searching is widespread. 
Of course, there has been considerable interest, 
especially as the result of Al research, in how to raake 
request formulation easier and more effective. Of greatest 
potential here are expert intermediaries that can function 
as front ends to an existing searching system and which 
enable the user to undertake good quality searches vdthout 
the knowledge or training demanded of the professional 
search intermediary. 
- 112 -
Given the coraplexity of the problem domain of 
online searching, it is evident that considerable 
expertise is reqired to enable Kood decisions to be made 
and the search to be conducted. There is not yet a 
detailed taxonomy of this expertise, although various 
writers on the design of front end systems have suggested 
categories which could be used to characterize this 
expertise. For example, Pollitt (1981) lists four 
categories: 
1 - system knowledge: the coramand language and 
facilities available in the search systera(s) frora logging 
on and the submission of search stateraents to the printing 
of references or abstracts; 
2 - searching knouledge: relating to the strategy and 
tactics to be employed in searching; 
3 - subject knouledge; 
4 - user knowledge: knowledge about each individual 
user including previous searches and preferred journals. 
It is added by Snith (1986): 
5 - database knovledge: fatniliarity with the content 
and structure of available databases. 
The fact is that more research is needed to 
determine in more detail the expertise underlying 
successful searching. 
A general review of expert intermediary systems 
is given by Marcus (1983). The main characteristic of 
these systems is that users are freed frcra encounters with 
- 113 -
the many peculiarities in databases and search svstems, 
and yet can benefit from a large range of capabilities, 
for example: 
users can enter a request in a loosely structured 
format, preferably in a natural language, sentence-like 
expression. An interraediary system processes the request 
tenns, displays information to the user (in the form of a 
list of subject areas, databases, search keys) from which 
users are asked to make a selection. Interactions of this 
nature usually proceed until users terminate the session; 
intermediary expert systeras can also replicate the 
performance of an expert in a particular area by 
incorporating the knowledge of an expert with rules for 
making inferences on the basis of this knowledge. 
The most interesting examples of front end 
systems are CONIT, EXPERT-1, and CANSEARCH which ali are 
being tested in experimental settings. 
CONIT (see Marcus, 1983) is a system that allous 
end users, who had no previous experience in operating 
retrieval systems, to obtain information, i.e., literature 
citations, from dozens of heterogeneous databases on four 
different host computer svstems. This system emphasizes a 
command/argument language structure to the interface. A 
development of CONIT, called EXPERT-1 seeks to simulate a 
human expert's search procedures in terms of search 
strategy formulation and explanation. The expertise built 
into EXPERT-1 includes the following abilities: 
- 114 -
- to assist the user to formulate his search problem; 
- to formulate a search strategy based on the concepts 
and search terms supplied by the user; 
- to assist the user to select the appropriate 
databases to search in; 
-to handle different protocols and command languages; 
- to assist the user reformulate search strategy based 
on partial search results. 
While CONIT - and most of the other intermediary 
systems - emphasizes a command/argument language approach 
to the user interface, EXPERT-1 employs only raenus and the 
fill-in-the-blank mode of computer - human interaction. 
A similar systems, CANSEARCH has been described 
by Pollitt (1984 ; see also 1986) to help doctors carry out 
online searches of cancer therapy in the MEDLINE database. 
Here again, the use of a human intermediary is replaced by 
a series of hierarchically organized menus that guide the 
user through the process of identifying those components 
of their problem that may need to be included in a query 
formulation. Thus, raenus are available for the 
specification of the site of the cancer, its type, the 
therapy that is under consideration, and the 
characteristics of the patient. Doctors can select the 
appropriate part of the menu on a touch-sensitive screen. 
When ali of the components of the search have been 
identified to the doctor's satisfaction, the system 
generates a Boolean auerv that is then submitted for 
processing by the MEDLINE system. The components of the 
- 115 -
CANSEARCH systera are shown in Fig. 8.: 
Rules 
Frame 
Selection 
State ment 
Generdtion 
System 
Interaction 
llockboarc/s. 
Thertify 
Site 
Type| 
Patient 
User 
Setetion 
Terminal Dnving Rules 
PROLOG 
Interpreter 
Utility Rules 
Svstem Dffving Rules 
Staternents 
^7 
Frarnes— 
COncepts and 
terms 
MEOLINE 
Fig. 8. Components of the CANSEARCH system (after 
Pollitt, 1984, p. 233). 
According to Town et al. (1985), such a design 
methodology would seem appropriate for end-user access to 
any reference database for which a well-designed thesaurus 
is available and for which the doraain of possible query 
types was sirailarly restricted. It is also interesting to 
stress, according to the same source, that the Commission 
of the European Communities has recently examined 
prcposals for a DIANE Intelligent Interface Facility 
- 116 -
(IIF) . The aim of IIF is to improve access to the 
databases available on the Euronet DIANE telecommunication 
network and, as far as possible, to make the network 
transparent to the user. 
Deriving from the relevance Al and expert 
svstems research in designing improved librarv/information 
svstems, it is in the librarv and information profession's 
interest to be aware of these developments. There is no 
doubt that the role of this profession is changing, for 
example, illustrated by Clarke and Cronin's statement 
(1983): 
"On the basis of the results of early trials there 
is a strong indication that the user of the future 
will not need the services of a librarian or 
information scientist in order to be able to 
conduct a coraprehensive and successful on-line 
literature search" (p. 286). 
Consequently, this raeans that the changes 
resulting from the application of Al research in 
library/in format ion services must also have an impact on 
library/information science education, i.e., it is the 
task of library/in format ion schools to prepare students 
for new roles and new careers in information transfer. 
- 117 -
CONCLUSIONS 
Throughout this discussion, the explosive growth 
of interest and work in Al and expert systems has been 
illustrated which, after optimistic beginning in the late 
1950's and a period of stagnation in the early 1970's, 
began again in the early 1980's. This is evident frora the 
Japanese Fifth Generation Computer Systems Project, the 
Alvey Programme in the U.K., ESPRIT, etc. One of the 
characteristics of this grovring interest is that, on the 
one hand, many governments in the West provided resources 
for researoh, and, on the other hand, commerce and 
industry shoved much greater openness to the ideas, 
techniques and tools of Al. The reason for this interest 
is, of course, the fact that Al is becorning a big, and 
increasingly growing business, most notably expressed in 
the areas of expert systems, natural languages, vision and 
robotics. 
Today, there is no doubt that Al has captured 
media interest and is fast becorning one of the most 
iraportant topics in discussions about the development of 
new technology. The indicators of this are the numerous 
articles in newspapers and journals, recently published 
books about this topic, newly established journals devoted 
only to this subject, etc. These published accounts relate 
to different disciplines, which emphasizes the position of 
Al as an interdisciplinary field. The reasons for the 
- 118 -
mul tidiscipl inarity of Al 1 ie, of course, in the term 
"intell igence" which is not well-defined and is, as such, 
included in many disciplines. 
In the context of the increasing interest in Al 
research, especially with regard to its commercial 
implications, there is one worrying feature: there are 
wide discrepancies in the reports about the achievements 
in the field. Throughout this study, raany examples have 
been quoted, one of the most alarming being PROSPECTOR, an 
expert system the aim of which is to aid geologists in 
evaluating mineral sites for potential deposits. On the 
one hand, this expert systera is the subject of many claims 
about the savings it had made for exploration companies, 
and, on the other hand, the designers of PROSPECTOR are 
impugnong such statements as untruth. 
Variations in the assessments of the state of 
the art were the starting-point for establishing a need 
for a closer association between Al and the social 
sciences, until the present only used by Al researchers to 
investigate the effects and impact of Al. If the reasons 
for disagreements about Al are to be understood, a social 
sciences approach - with sociology taking the central role 
- to the discourse of Al has to be applied. 
It has been str-esed in one sections that the 
reasons for discrepancies in reports about Al and expert 
systems lie in the interpretative flexibility of the term 
"intelligence". What counts as "intelligence" and the 
questions about whether machines can be intelligent or not 
- 119 -
are at the heart of these discrepancies. For exaraple, 
optimistio views, representations of the achievements of 
the field might be expected from those involved in 
marketing Al applications. However, these reports are 
elsewhere (in the less "popular" press) countered by 
considerable caution and pesisimism about the achievements 
to date. To recognize the interpretative flexibility of 
this idea of "intelligence" is a very important goal for 
the social sciences because otherwise they will have to 
wait for the output of Al research, rather than be 
involved in a detailed consideration of the process of the 
research activity itself. 
It follows that the social sciences should not 
only be concerned with the effects of Al, but also with 
its genesis. The social sciences will only be capable of 
assessing realistically the irapact and effects of Al 
research if they also have insight into the key issues of 
Al, one of the most important being knowledge 
representation. There is no doubt that knovledge is also a 
social concept, which again indicates the need for a 
closer relationship between Al research (e.g., knowledge 
engineering) and the social sciences. 
For many years it has been thought that the 
development of cognitive psychology and the so-called 
"coraputational metaphor" is the only relationship between 
social sciences and Al. In addition, many introductory 
3ections of Al literature also include citations about the 
relevance of psychology and linguistics to Al. What is 
- 120 -
surprising is the fact that the important possibility of 
an association between sociology and Al has hardly been 
noticed. There is no doubt that sociology, whose interests 
clearly encompass language use and interaction, 
intentionality, knouledge systeras, etc., is as yet 
unexplored territory. And, perhaps even more important, 
knowledge is recognized in the modern sociology of 
knowledge as a legitimate sociological object. Therefore, 
a two-sided exclusion of sociology from Al research has 
been discussed: 
1 - Al researchers have not been interested in the 
possible contributions of sociological research in, for 
example, knowledge systems; 
2 - the function of sociology has been reduced to the 
investigations about the impact and effects of Al, instead 
on the Al research activity itself as a condition for such 
investigations. 
It has been shown that this reduction of 
sociological capability corresponds to the pre-Kuhnian 
view of science, where "social factors" were precisely 
those factors not related to "science itself"; the dcmain 
of the "social" was regarded as outside or (at best) 
peripheral to the actual science. But as the post-Kuhnian 
sociology has established the nature and content of 
scientific knowledge as legitimate object, the task of 
sociology is to break the barrier between "the social" and 
"the scientific" in the context of Al research. 
- 121 -
In the second part of this study, the assumption 
that the social sciences should also be concerned with the 
genesis of Al was tested by the example of expert systems. 
Although expert systems are recognized as the 
applied end of Al research, there are many controversies 
in the field. Extraordinary optimisra is very often 
countered with reports that the area faces fundamental 
probleras. An additional problem is a lack of agreement in 
expert systeras literature about even the basic issues such 
as definition, aims, essential characteristics, 
evaluation, commercial use, etc, of expert systems. 
Moreover, many over-siraplified descriptions of the field 
can also be found. Such an "atheoretical" foundation 
leads, of course, to difficulties when trying to assess 
the future development of expert systems, especially with 
regard to their their possible effects and impact in 
different environraents, issues which are usually addressed 
to the social sciences. 
From this starting-point the question as to what 
was irapeding greater clearness in expert systems 
discussions was raised. It has been stressed that only a 
detailed analysis of crucial expert svstems issues, such 
as knowledge acquisition, knowledge representation, and 
explanation facilities can provide the answer to this 
question. There is no doubt that these issues are also 
social concepts and, as such, relevant for the social 
sciences. This notion is very important because, on the 
one hand, expert systeras will be able to meet the 
- 122 -
challenge of general competence and reliability if more 
fundamental progress can be made by the social sciences 
(e.g., psychology, sociology of knowledge, etc.) in 
understanding the structure of knowledge and the whole 
knowledge complex, i .e., transfer of knowledge, 
formalization of knowledge, process of reasoning, common 
sense, etc.; and, on the other hand, this understanding is 
a starting-point to see where the development of expert 
systems is leading. Only such an established framework 
will enable the social sciences to discuss realistically 
problems of the impact and effects of expert systems 
(e.g., how the status of the human experts will be 
affected by expert systems, legal implications of the use 
of expert systems, etc.). 
Throughout this study the notion that knouledge 
representation is a vital issue in Al research today has 
been followed, i.e., its goal is to provide tools that 
exploit new ways to encode and use knowledge to solve 
problems. Without regard to some fundamental problems 
connected with knowledge representation schemes, such 
research results can be important in many different 
environments, including the design of improved 
library/information systems. Two areas where such research 
can be useful for library/information services have been 
identified: 
- providing new methods for crganizing and 
representing information in databases, i.e., "system 
thinking" whicn means an organized way of linking bits of 
- 123 -
information into networks, shouring the relationships 
between data; 
- developing expert intermediary systeras as front ends 
to bibliographic databases. 
In this connection it has also been emphasized 
that these changes resulting from Al and expert systems 
research should have an irapact on library/information 
education. 
Finally, it is appropriate once more to stress 
that when applying new methods of organizing information 
and knowledge in databases it is necessary to put the 
structure of knowledge in the first plače, i.e., it is 
necessary to recognize a distinction between fields where 
knowledge can be formalized in a highly structured way 
(chemistry, mathematics) and fields with "weak" forraalism 
(social sciences). In this light - apart from the 
unjustified distinction between information services and 
libraries - the research project "Development of 
scientific and technical information in Slovenja 1986-90"i 
mentioned in the introductory section, should be 
considered. Its relevance is in stressing the need for the 
development of new methods of organizing information and 
knovledge in database systems, but in doing that it should 
also take into- account different relationships between 
knowledge, communication, and information systems in the 
sciences, the social sciences, and the humanties. 
- 124 -
NOTES 
1. These differences can be illustrated by the example of 
chemical information systems where the techniques used 
fall into two categories: firstly, those which involve 
the analysis and organization of text; secondly, those 
which are concerned with handling chemical structural 
information. Chemical data is unusual because the 
second aspect of it, the structure of a molecule, 
cannot be handled by normal bibliographic methods. For 
more information on this topic see Ash and Hyde 
(1975). 
2. According to Becker (1986), Eugene Charniak and Drew 
McDermott trace in their book "Introduction to 
Artificial Intelligence" the first use of the 
terminology. They explain that in 1956 John McCarthy, 
an assistant professor of mathematics at Dartmouth 
College, and Marvin Minsky from MIT organised a 
conference in Dartmouth College in New Hampshire. 
During the conference McCarthy proposed that a study 
of Al should be carried out at the college to proceed 
on the basis of the conjecture that every aspect of 
learning, or any other feature of intelligence, can in 
principle be so precisely described that a machine can 
be made to simulate it. 
3. Fundamental aspects of Al are clearly presented in 
books by Nilsson (1982), and Barr and Feigenbaum 
(1981). 
4. Different positions to Al research are concisely 
described by Fleck (1984). 
5. More about the construct "intentionality" can be found 
in Searle (1983). 
6. This conversation is taken from Weizenbaum (1985). 
7. Ed Vfeizenbaum (1985) illustrated the dangers of work 
in Al by one very interesting comment on ELIZA, 
written by an enthusiastic psychotherapist: "Further 
work most be done before the program will be ready for 
clinical use. If the methods proves beneficial, then 
it would provide a therapeutic tool which can be made 
- 125 -
widely available to mental hospitals and psychiatric 
centres suffering a shortage of therapists. Because of 
the time-sharing capabilities of modern future 
computers, several hundred patients on hour could be 
handled by a computer syste^i designed for this 
purpose. The human therapist, involved in the design 
and operation of this system, wouid not be replaced, 
but would become a much more efficient man since his 
efforts would no longer be limited to the one-to-one 
patient-therapist ratio as now exists" (p. 5). 
8, LISP (short for LISt Processing) was developed by J. 
McCarthy and his associates at MIT during the late 
1950s and early 1960s. 
9. The programme is naraed after Mr. John Alvey of British 
Telecom, chairman of the committee which in 1982 
recommended that such a national programme should be 
mounted, in response to increasing overseas 
competition and in particular to the Japanese Fifth 
Generation Computer Systems initiative. 
10. Some details about the Alvey Programme are taken from 
its Annual Report 1985. 
11. For example, the "Univision" system used by Unimation 
based on a vision systera developed from Al research 
for the market by MIC, Machine Intelligence 
Corporation (see Winston and Prendergast, 1984). 
12. The first report on PROLOG (PROgramming in LOGic) was 
first published in 1975 by researchers based at the 
University of Marseilles. 
13- The examples of the dialogue with MYCIN are taken from 
Daviš (1984). 
14. There are also two alternative methods, i.e., 
propagation of constraints and problem reduction which 
have already been described in the section on the 
subareas of Al. 
15. Knowledge representation scheraes are succinctly 
described by Mylopoulos and Levesque (1983) > and in 
three articles, edited by Friedland (1985). 
- 126 -
16. Exaraples for PROLOG are taken from Michie and Johnston 
(1985). 
17. The idea of "tacit" knowledge is taken from M. Polany 
(1976). Polany's well known example of tacit knowledge 
is the skill associated with bicycle riding. The 
formal dynamics of balance on a bicycle riding do not 
comprise the rules of riding. A rider may know nothing 
of centres of gravity and gyroscopic forces yet stili 
rides whereas the most expert bicycle engineers may 
not be able to do so. The rider knows how to ride 
without being able to say hov/. 
- 127 -
BIBLIOGRAPHY 
Adams, J.B. (1976) 'A probability model of medical 
reasoning and the MYCIN model.' Mathematical Biosciences, 
32, 177-186. 
Addis, T.R. (1982) 'Expert systems: an evolution in 
information retrieval. ' Information Technology: Research 
and Development, _1, 301-324. 
Adler, F.R. (1984) An investment opportunity? I_n: Winston, 
P.H., and Prendergast, K.A. (eds.) The Al business: the 
commercial ušes of artificial intelligence. Cambridge, 
Mass : MIT, pp. 255-262. 
Aleksander, I. (1984) Designing intelligent systems: an 
introduction. London : Kogan Page. 
Alexander, T. (1982) 'Practical ušes for a "useless" 
science.' Fortune, 105, 138-145. 
Alvey Programme, Annual Report 1985. London : Alvey 
Directorate. 
Ash, J.E., Hayde, E. (eds.) (1975) Chemical information 
systems. Chichester : Ellis Horwood. 
Barr, A., Feigenbaum, E.A. (eds.) (1981) The handbook of 
artificial intelligence. Vol. 1. London : Pitman. 
Bateman, J. (1985) The role of language in the raaintenance 
of intersubjectivity: a computational investigation. in: 
Gilbert, G.N., and Heath, C. (eds.) Social action and 
artificial intelligence. Aldershot : Gower Press, pp. 40-
Becker, J. (1986) 'On good with words.' Expert Systems 
User, _1, 20-21. 
Belkin, N.J., Vickery, A. (1935) Interaction in 
information systems: a review of research from document 
retrieval to knowledge-based systems. London : British 
Library. 
- 128 -
Bernstein, L.M., Siegel , E.R., Goldstein, CM. (1980) 'The 
hepatitis knowledge base: a prototype information transfer 
system.' Annals of Internal Medicine, 43, 165-222. 
Bivins, K.T., Palmer, R.C. (1981) 'A microcomputer 
alternative for information handling: REFLES.' Information 
Processing and Management, 17, 93-101. 
Bobrov/, D. G., Winograd , T. (1977) 'An overview of KRL, a 
knowledge representation languase.' Cognitive Science, _1, 
3-46. 
Boden, M. (1977) Artificial intelligence and natural man. 
Hassocks : Harvester Press. 
Boden, M. (1984) 'Artificial intelligence and social 
forecasting.' Journal of Mathematical Sociology, _9, 341-
356. ' "*"" 
Borko, H. (1985) 'Artificial intelligence and expert 
systems research and their possible impact on information 
science education.' Education for Information, _3, 103-114. 
Bramer, M.A. (1981) A survey and critical reviev/ of expert 
systems research. _In: Parslow, R.D. (ed.) Information 
Technology in the Eighties. London : Heyden. 
Bramer, M.A. (1985) Expert systeras: the vision and the 
reality. jln: Bramer, M.A. (ed.) Research and development 
in expert systems. Carabridge : Cambridge University Press, 
pp. 1-12. 
Buchanan, B.G., Duda, R.O. (1983) 'Principles of rule-
based systems.' Advances in Computers, 22, 163-216. 
Buchanan, B.G., Feigenbaum, E.A. (1978) 'DENDRAL and META-
DENDRAL: their applications dimensions.' Artificial 
Intelligence, 11, 5-24. 
Buchanan et al. (1 983 ) Constructing an expert system . In : 
Hayes-Roth, F., Waterman, D., Lenat, D.B. (eds.) Building 
expert systems. Reading, MA : Addison-Wesley, pp. 127-167. 
- 129 -
Carhart, R.E. (1979) CONGEN: an expert system aiding the 
structural chemist. In: Michie, D. (ed.) Expert systems in 
the micro-electronic age. Edinburgh : Edinburgh University 
Press , pp . 65-82. 
Cendrowska, J., Bramer, M.A. (1984) 'A rational 
reconstruction of the MYCIN consultation system.' 
International Journal of Man-Machine Studies, 20, 229-317. 
Cercone, N. , McCalla, G. (1984) »Artificial intelligence: 
underlying assumptions and basic objectives.' Journal of 
the American Society for Information Science, 35~j 280-2 90. 
Clarke, A., Cronin, B. (1983) 'Expert systems and 
library/information work.' Journal of Librarianship, 1 5, 
2 77-2 92. 
Clocksin, W.F., Mellish, C.S. (1981) Programming in 
PROLOG, New York : Springer-Verlag. 
Collins, H.M. (1986) The concept of explanation in expert 
systems. (Paper presented at conference on "Explanation in 
Expert Systems" , University of Surrey, March 20-21, 1986). 
Collins, H.M., Green, R.H., Draper, R.C. (1985) Where's 
the expertise? : expert systeras as a medium of knowledge 
transfer. Jn_: Merry, M. J. (ed.) Expert systems 85. 
Cambridge : Cambridge University Press, pp. 323-334. 
Coulter, J. (1985) On comprehension and "mental 
representation". ^H: Gilbert, G.N., Heath, C. (eds.) 
Social action and artificial intelligence. Aldershot : 
Gower Press, pp. 8-23. 
D'Agapeyeff, A. (1984) Report to the Alvey Directorate on 
a short survey of expert systems in UK business. London : 
Alvey Directorate. 
Daviš, R. (1982) 'Expert systeras: Where are we? And where 
do we go from here?' The Al Magazine, _3_, 3-22. 
Daviš, R. (1984) Amplifying expertise with expert systems. 
In: Winston, P. H. , Prendergast, K.A. (eds.) The Al 
business: the commercial ušes of artificial intelligence. 
Cambridge, Mass : MIT, pp~! 17-40. 
- 130 -
Daviš, R., Buchanan, B.G. , Shortliffe, E.H. (1977) 
'Production rules as a representation for a knowledge-
based consultation program.' Artificial Intelligence, _8, 
15-45-
Davies, R., James, B. (1984) 'Towards an expert svstem for 
cataloguing: some experiments based on AACR2.' Program, 
J_8, 283-297. 
Dennett, D. (1979) Brainstorms: philosophical essays on 
mind and psychology. Hassocks : Harvester Press. 
Denning, P. J. (1986) 'The science of computing: expert 
systems.' American Scientist, 74, 18-20. 
Duda, R.O., Gaschnig, J.G. (1981) 'Knowledge-based expert 
systems come of age.' Byte, j>, 238-281. 
Duda, R.O., Hart, P.E., Reboh, R. (1985) 'Letter to the 
editor.' Artificial Intelligence, 26, 359-360. 
Duda, R.O., Shortliffe, E.H. (1983) 'Expert systems 
research.' Science, 220, 261-268. 
Ernst, G.W., Newell, A. (1969) GPS: A čase study in 
generality and problem solving. New York : Academic Press. 
Evanczuk, S., Manuel, T. (1983) 'Practical systems use 
natural languages and store human expertise.' Electronics, 
December 1, 139-145. 
Feigenbaum, E.A. (1979) Themes and čase studies of 
knouledge engineering. _In: Michie, D. (ed.) Expert systems 
in the micro-electronic age. Edinburgh : Edinburgh 
University Press, pp. 3-25. 
Feigenbaum, E.A., McCorduck, P. (1984) The Fifth 
Generation: artificial intelligence and Japan's computer 
challenge to the world. London : Michael Joseph. 
Fikes, R., Kehler, T. (1985) 'The role of frame-based 
representation in reasoning.' Communications of the ACM, 
2_8, 904-920. 
- 131 -
Fleck, J. (1982) Development and establishment in 
artificial intelligence. Jn_: Elias, N. , Martins, H., 
Whitley, R. (eds.) Scientific Establishments and 
Hierarchies. Sociology of the Sciences Yearbook, Dordrecht 
: Reidel, 6, 169-217T 
Fleck, J. (1984) Artificial intelligence and industrial 
robots: an autoraatic end for utopian thought In: 
Mendelsohn, E. , Nowotny, H. (eds.) Nineteen Eighty-Four: 
Science Between Utopia and Dystopia. Sociology of the 
Sciences Yearbook, Dordrecht : Reidel , _8, 189-2317 
Friedland, P. (ed.) (1985) 'Knowledge-based 
architectures.' Communications of the ACM, 28. 
Genesereth, M.R., Ginsberg, M.L. (1985) 'Logic 
programraing.' Communications of the ACM, 28, 933-941. 
Gilbert, G.N., Heath, C. (eds.) (1985) Social action and 
artificial intelligence. Aldershot : Gower Press. 
Goldstein, I., Papert, S. (1977) 'Artificial intelligence, 
language and the study of knowledge.' Cognitive Science, 
]_. 84-123. 
Gooding, C. (1986) 'Only huraans can be expert.' Computer 
Weekly, May 8, p. 39. 
Guilfoyle, Ch. (1986a) 'A table load full of micro 
shells.' Expert Systetns User, 2, 18-20. 
Guilfoyle, Ch. (1986b) 'Expert explanation.' Expert 
Systems User, 2, 25-27. 
Hayes-Roth, F. (1985a) 'Knowledge-based expert systems -
the state of the art in the US.' Knowledge Engineering 
Review, J, 18-27. 
Hayes-Roth, F. (1985b) 'Rule-based systems.' 
Communications of the ACM, 2_8, 921-932. 
Hayes-Roth, F., Waterman, D.A., Lenat, D.B. (1983) An 
overview of expert sy3tems. _IH: Hayes-Roth, F., Waterman, 
D., Lenat, D.B. (eds.) Building expert systems. Reading, 
MA : Addison-Wesley, pp. 3-29. 
- 132 -
Hjerppe, R. (1983) What artificial intelligence can, 
could, and can11 do for libraries and information 
services. Proceedings of the 7th IOLIM, Dec 6-8, 1983, 
London : Learned Information, pp. 7-25. 
Hjerppe, R., Olander, B., Marklund, K. (1985) Project 
ESSCAPE - Expert Systeras for Simple Choice of Access 
Points for Entries: applications of artificial 
intelligence in cataloguing. (Paper presented at IFLA 51st 
General Conference in Chicago, 18—2M August, 1985). 
Johnston, R. (1986) 'ART and KEE: are they worth the 
priče?' Expert Systems User, _1_> 10-11. 
Kornhauser, A. (1983) Teaching chemistry today: what do we 
aim for? _In_: Grunevald, H. (ed.) Chemistry for the future. 
0xford : Pergamon Press, pp. 383-392. 
Kornhauser, A. (1985) Razvoj v smeri inženiringa znanja 
(Towards knowledge engineering) . (1. posvetovanje Sekcije 
za specialne knjižnice Zveze bibliotekarskih društev 
Slovenije - Paper presented at the 1st Conference of the 
Section for Special Libraries of the Library Association 
of Slovenia, Ljubljana, 12 November, 1985). 
Kowalski, R.A. (1977) Predicate logic as a programming 
language. New York : North-Holland. 
Kraf t, A. (1 984 ) XC0N: an expert configuration system at 
Digital Equipment Corporation. _Jii: Winston, P.H., 
Prendergast, K.A. (eds.) The Al business: the commercial 
ušes of artificial intelligence. Cambridge, Mass : MIT, 
pp. 41-49. 
Kuhn, T. S. (1962) The structure of scientific revolutions. 
Chicago : University of Chicago Press. 
Leith, P. (1986) ' Why the experts must be accountable.' 
Computer Guardian, May 8, p. 15. 
Lenat, D.B. (1977) On automated scientific theory 
formation: a čase study using the AM program. In: Hayes, 
J.E. , Michie, D., Mickulich, L.I. (edsTT Machine 
Intelligence 9- New York : Halsted, pp. 251-2 86. 
- 133 -
Lenat, D.B. (1982) 'The nature of heuristics.' Artificial 
Intelligence, 19, 189-221. 
Lenat, D.B. (198*0 'Computer software for intelligent 
systems.' Scientific American, 251, 152-160. 
Lindsay et al. (1980) Applications of artificial 
intelligence for organic chemistry: The DENDRAL project. 
New York : McGraw-Hill. 
McDermott, J. (1982) 'R1: a rule-based configurer of 
computer systems.' Artificial Intelligence, 19, 39-88. 
Manchester, P. (1986) 'Clubs take over cult of artificial 
intelligence.' Computer Weekly, May 8, p. 38. 
Marcus, R. S. (1983) 'An experimental comparison of the 
effectiveness of computers and humans as search 
intermediaries.' Journal of the American Society for 
Information Science, 34, 381-40*1. 
Michard, A. (1982) 'Graphical presentation of boolean 
expressions in a database query language.' Behaviour and 
Information Technology, J, 279-288. 
Michie, D. (1984) 'Automating the synthesis of expert 
knowledge.' Aslib Proceedings, 36, 337-343-
Michie, D., Johnston, R. (1985) The creative computer: 
machine intelligence and human knovdedge. Harraondsvrorth : 
Penguin. 
Minsky, M.L. (1979) Computer science and the 
re presentation of knowledge. JEn: Dertouzos, M. L., Moses, 
J. (eds.) The computer age: a twenty-year view. Cambridge, 
Mass : MIT, pp. 392-421. 
Mozetič, I., Brat ko, I., Lavrac, N. (1983) An experiment 
in automatic synthesis of expert knowledge through 
qualitative modelling. Proč. Logic Prog. Vforkshop, 
Albufeira, Portugal, June 26 - July 1, 1983. 
- 134 -
Mylopoulos, J., Levesque, H. J. (1983) An overview of 
knowledge representation. _In_: Brodie, M., Mylopoulos, J., 
Scnmidt, J. (eds.) On conceptual modelling. New York : 
Springer-Verlag , pp . 3-17. 
Nau, D.S. (1983) 'Expert computer systems.' Computer, 16, 
63-85. 
Newell, A. (1983) Intellectual issues in the history of 
artificial intelligence. _In: Machlup, F., Mansfield, U. 
(eds.) The study of information. New York : Wiley, pp. 
187-227. 
Newell, A., Simon, H.A. (1972) Human problem solving. 
Englevrood Cliffs, N.J. : Prentice-Hall. 
Nilsson, N.J. (1971) Problem-solving methods in artificial 
intelligence. New York : McGraw-Hill. 
Nilsson, N.J. (1982) Principles of artificial 
intelligence. Berlin : Springer. 
Partridge, D. (1986) Social implications of artificial 
intelligence. In: Yazdani, M. (ed.) Artificial 
intelligence: principles and applications. New York : 
Chapman and Hali, pp. 315-336. 
Pogson, B. (1986) 'Emulated.' Computer Guardian, May 22, 
p. 17. 
Polanvi, M. (1967) The tacit dimension. New York : Anchor. 
Pollitt, A.S. (1981) An expert system as an online search 
intermediary. Jjn: Proceedings of the 5th International 
Online Information Meeting. Oxford : Learned Information, 
pp. 25-32. 
Pollitt, A. S. (1984) 'A "front-end" system: an expert 
system as an online search intermediary. ' Aslib 
Proceedings, _36, 229-234. 
Pollitt, A.S. (1986) An expert systems approach to 
document retrieval: a summary of the CANSEARCH Research 
Project. Huddersfield Polytechnic, (Technical Report 
Series, no 86/6 ). 
- 135 -
Quinlan, J. R. (1979) Discovering rules by induction from 
large collections of examples. I_n_: Michie, D. (ed .) Expert 
systems in the micro-electronic age. Edinburgh : Edinburgh 
University Press, pp. 168-201. 
Ritchie, G.D., Hanna, F.K. (1982) AM: a čase study in Al 
methodology. Edinburgh : University of Edinburgh 
(Department of Artificial Intelligence, Technical Report 
No. TR18). 
Robinson, J. (1965) 'A machine oriented logic based on the 
resolution principle.' Journal of the ACM, 12, 23-41. 
Schank, R.C. (1975) Conceptual information processing. New 
York : North Holland. 
Searle, J. (1980) 'Minds, brains and programs 
Behavioural and Brain Sciences, 2, 417-457. 
Searle, J. (1983) Intentionality: an essay in the 
philosophy of mind. Cambridge : Cambridge University 
Press. 
Sheil, B. (1983) 'Power tools for programmers.' 
Datamation, February, 131 —1U4. 
Shortliffe, E.H. (1976) Computer based medical 
consultations: MYCIN. New York : American Elsevier. 
Small, H. (1984) The lives of a scientific paper. In: 
Warren, K. (ed.) Selectivity in information systems. New 
York : Praeger, pp. 83-97. 
Small, H. (1986) 'The svnthesis of specialty narratives 
from co-citation clusters.' Journal of the American 
Society ;for Information Science, 37, 97-110. 
Smith, L.C. (1976) 'Artificial intelligence in information 
retrieval systems.' Information Processing and Management, 
12, 189-222. ~ "~ ~" 
Smith, L.C. (1985) Knovledge-based systems, artificial 
intelligence and human factors. (Paper prepared for the 
Seminar on Information Technology as a Tool for' 
Information Use, Copenhagen, 8-10 May, 1985). 
- 136 -
Smith, L.C. (1 986) Machine intelligence vs. machine-aided 
intelligence as a basis for interface dasign. (Paper 
presented at 2nd Conference on Computer Interfaces and 
Intermediaries for Information Retrieval, 29 May, Boston, 
1986). 
Sowizral, H. A. (1985) 'Expert systems.' Annual Review of 
Information Science and Technology, 20, 179-199-
Stamper, R. (1985) Knowledge as action: a logic of social 
norms and individual affordances. ^n: Gilbert, G.N. , 
Heath, C. (eds.) Social action and artificial 
intelligence. Aldershot : Gower Press, pp. 172-191. 
Stefik, M. (1981) 'Planning with constraints (MOLGEN Part 
1).' Artificial Intelligence, J_6, 111-140. 
Stefik, M. (1981) 'Planning and meta-planning (MOLGEN Part 
2).' Artificial Intelligence, J_6, 141-170. 
Town, W.G. et al. (1985) Advanced computational support 
for biotechnology research: project report. Prepared by 
Hampden Data Services, U.K., on behalf of Coramission of 
the European Comraunities ITTTF/CUBE, May 1985-
Turing, A. (1950) 'Computing machinery and intelligence.* 
Mind, 5_9, 433-460. 
van Melle, W. (1980) A domain-independent system that aids 
in constructing knowledge-based consultation programs. 
Stanford Heuristic Programming Project Merao HPP-80-22. 
Vickery, B.C. (1968) 'Bibliographic description, 
arrangement and retrieval.' Journal of Documentation, 24, 
1-15. 
Waldrop, M.M. (1984) 'Artificial intelligence (I): into 
the world.' Science, 22J, 802-805. 
Walker, D.E. (1981) 'The organization and use of 
information: contributions of information science, 
computational linguistics and artificial intelligence.' 
Journal of the American Society for Information Science, 
32, 347-363. 
- 137 -
Waltz, D.L. (1982) 'Artificial intelligence.' Scientific 
American, 2kl, 101-122. 
Ward , S. A., Reed , L. J. (eds.) (1983) Knowledge structure 
and use: implications for synthesis and interpretation. 
Philadelphia : Teraple University Press. 
VJebster, R. , Miner, L. (1982) 'Expert systems: programming 
problem-solving .' Technology, _2, 62-73. 
Weizenbaum, J. (1985) Computer power and human reason: 
from judgment to calculation. Harmondsworth : Penguin. 
Vfelbank, M. (1983) A revievr of knowledge acguisition 
technigues for expert systems. Martieshara Heath : British 
Telecom Research Laboratories. 
Winograd, T. (1972) Understanding natural language. New 
York : Academic Press. 
Winston, P.H., Prendergast, K.A. (eds.) (1984) The Al 
business: the commercial ušes of artificial intelligence. 
Cambridge, Mass : MIT. 
Vfoolgar, S. (1985) 'Why not a sociology of machines? The 
čase of sociology and artificial intelligence.' Sociology, 
19, 557-572. 
Yaghmai, N.S., Maxin, J. A. (1984) 'Expert systems: a 
tutorial.' Journal of the American Society for Information 
Science, J5~ 297-305. 
Zadeh, L.A. (1979) A theory of approximate reasoning. In: 
Hayes, J.E., Michie, D., Mickulich, L. I. (eds.) Machine 
Intelligence 9. New York : Wiley. 
- 138 -