Book Reviews 
Associative Networks - 
Representation and Use 
of Knowledge by Computers 
Nicholas V. Findler, Editor 
Academic Press, New York, 1979, 
462 pp., $42.50, ISBN 0-12-256380-8. 
Upon opening this book and leafing through the 
pages, one gets the impression of an important com- 
pendium. The fourteen articles provide good coverage 
of semantic networks and related systems for repre- 
senting knowledge. Their average length of 33 pages 
is long enough to give each author reasonable scope, 
yet short enough to permit a variety of viewpoints to 
be expressed in a single volume. The editor should be 
commended for his efforts in putting together a well- 
organized book instead of just another collection of 
unrelated papers. 
In the first article, On the Epistemological Status of 
Semantic Networks, Ronald Brachman surveys semantic 
networks since 1966 and introduces many of the issues 
that are discussed by other authors in later chapters. 
His main point is the need for greater structure in the 
nets than a uniform graph of concept nodes and rela- 
tionship arcs. In his system, KLONE, generic con- 
cepts are themselves defined by a network, which is 
inherited either by individual concepts that are in- 
stances of the generic or by other generic concepts 
that represent subtypes of the higher level generic. 
As a survey, Brachman's article is good; but as a 
history, it is flawed by the common error that semantic 
networks "originated in the work of Quillian." Yet 
semantic networks, like much of the AI work on natu- 
ral language, are a re-invention of what was done for 
machine translation a decade earlier. In linguistics, the 
most extensive network representation was the depen- 
dency grammar by Lucien TesniCre (1959). Although 
TesniCre never used a computer himself, he had a 
strong influence on MT work in Europe. In the Unit- 
ed States, dependency parsers were implemented by 
Hays (1964) and Klein (1965), whom Schank and 
Tesler (1969) cited as an influence on their work. Of 
the semantic networks for MT, the most sophisticated 
were the correlational nets by Silvio Ceccato (1956, 
1961, 1962). His work was so advanced that Bar- 
Hillel (1960) dismissed it as "practically hopeless" 
because of its obscure speculations about "elements of 
thought." Contrary to Brachman's claim that 
Carbonell's nets were the first to distinguish subtypes 
from instances of a type, Ceccato had distinct rela- 
tions for member-class, species-genus, and part-whole. 
Other arcs in Ceccato's nets represented linguistic 
cases such as object, instrument, and result and spatial 
relations such as containing, covering, and contiguity. 
Ceceato implemented the typical AI mechanisms of 
inheritance of properties from genus to species and 
frame-like "constellations" which "indicate both the 
particular relations that extend from the analyzed 
thing, and the things that can satisfy these relations" 
(1961, p. 65). Given the concepts PAINTBRUSH and 
PICTURE, he could trace a path through a constella- 
tion to show that PAINTBRUSH was the instrument 
of PAINT and PICTURE was the result. 
Two other articles in the book introduce additional 
structure into semantic networks. In A Procedural 
Semantics for Semantic Networks, Hector Levesque 
and John Mylopoulos develop inheritance mechanisms 
similar to Brachman's, but with a different notation. 
Their main innovation is the merger of semantic net- 
works with a network notation \[or procedures and 
PLANNER-like conditions for invoking the proce- 
dures. Their work forms a bridge between declarative 
forms of semantic nets and procedural languages for 
knowledge representation. In Making Preferences 
More Active, Yorick Wilks extends his earlier work on 
preference semantics with more frame-like structures 
called pseudotexts. He starts with an example from the 
London Times and shows how his system would deal 
with typical violations of selectional constraints. His 
analysis is worth reading by anyone who is program- 
ming computers to understand natural language, al- 
though it is unlikely that others will be quick to adopt 
Wilks' exuberant terminology of formulas, templates, 
paraplates, semantic blocks, extractions, and pseudo- 
texts. 
Several articles develop notations for representing 
quantifiers and specifying their scope. In Encoding 
Knowledge in Partitioned Networks, Gary Hendrix pres- 
ents a general method of nesting subnets inside of 
other nets. The nesting shows scope of quantifiers, 
but it could also support other uses for nested contexts 
such as modalities. Hendrix assumes that the occur- 
rence of a concept automatically asserts its existence; 
but if a generic concept occurs in both the antecedent 
and the consequent of an implication, it is assumed to 
have a universal quantifier. In Extensional Semantic 
Networks, Jtirgen Janas and Camilla Schwind also use 
partitioning to show scope of quantifiers, but they 
append explicit existential or universal quantifiers to 
the generic concepts. Despite the title of their article, 
the authors do represent intensional information in the 
networks and present a good discussion of the distinc- 
tion between intensions and extensions. In The Struc- 
ture and Organization of a Semantic Net for Compre- 
hension and Inference, Lenhart Schubert, Randolph 
Goebel, and Nicholas Cercone introduce various kinds 
of dotted, dashed, and double links that show scope of 
Boolean operators, quantifiers, and modality. Their 
notation is less readable than Hendrix's or Janas and 
Schwind's, but they discuss general issues of organiz- 
110 American Journal of Computational Linguistics, Volume 6, Number 2, April-June 1980 
Book Reviews Associative Networks - Representation and Use of Knowledge by Computers 
ing and retrieving knowledge that would apply to al- 
most any representation. In The SNePS Semantic 
Network Processing System, Stuart Shapiro presents an 
extensional approach to handling quantifiers with 
structure-building rules; "Every man loves some wom- 
an," for example, corresponds to a rule that builds a 
structure for "x loves a woman y" whenever it finds a 
node for a man x. That approach works when all of 
the facts are represented in a single context. But to 
represent "Tom believes that every man loves some 
woman" would require a strange nesting of structure- 
building rules inside the context of the verb comple- 
ment of BELIEVE. 
In A Predicate Calculus Based Semantic Network for 
Deductive Searching, James MeSkimin and Jack Minker 
present a solid, well-written article on sorted first- 
order predicate calculus for data base inference. But 
calling their notation a semantic network is highly 
misleading. According to Brachman, the common 
thread that holds together the various forms of seman- 
tic networks is their use "in understanding natural 
language"; and according to Janas and Schwind, 
"intensionality is undoubtedly a constituent of any 
semantic network." McSkimin and Minker's approach 
fails both of these criteria: they make no attempt to 
deal with natural language, and their semantic categor- 
ies are "simply labels for a set of objects." With their 
purely extensional approach, the concept UNICORN 
would be a subconcept of anything because the set of 
all unicorns, being empty, is a subset of all other sets. 
Three of the articles apply some form of semantic 
networks to specific problems in computational lin- 
guistics. In Re: The Gettysburg Address, Roger Schank 
and Jaime Carbonell, Jr., analyze Lincoln's famous 
prose as a stimulating challenge for conceptual depen- 
dency (CD) theory. As a result of the analysis, they 
extend the list of primitive CD ACTs with seven social 
ACTs such as AUTHORIZE, ORDER, and PETI- 
TION. The social ACTs are high-level concepts that 
may optionally be expanded in terms of the low-level 
primitives. Introducing them solves a common prob- 
lem of CD theory: too much expansion of concept 
nodes can lead to a combinatorial explosion of primi- 
tive nodes. In Rule Forms for Verse, Sentences, and 
Story Trees, Robert Simmons and Alfred Correira draw 
some parallels between story generation, problem solv- 
ing, and theorem proving. As examples, they cite 
Meehan's Metanovel as well as some of their own 
work. Although Meehan's underlying representation 
uses conceptual dependency graphs instead of standard 
predicate calculus, the output of his story generator 
may be viewed as a trace of the steps taken by a clas- 
sical theorem prover. In On Representing Common- 
sense Knowledge, Benjamin Kuipers gives a working 
definition of commonsense knowledge, discusses some 
constraints on systems that have it, and then proposes 
a representation based on a partial order and associa- 
tive triples. His definition seems acceptable, but the 
proposed representation is one of the least structured 
in the book, and the example he gives is too limited to 
show that the representation can support the defini- 
tion. 
Three other articles discuss systems that use some 
form of semantic networks. In A Heuristic Information 
Retrieval System Based on Associative Networks, Nicho- 
las Findler mentions a keyword, ELIZA-like system 
that he developed, proposes a new information retriev- 
al system called IRUHS-1, describes some clustering 
techniques for fuzzy retrieval, and then gives a sample 
dialog that he would like IRUHS-1 to support. Yet 
the dialog would require sophisticated syntax, seman- 
tics, and deductive capabilities that the body of the 
article never discusses. In Representations to Aid Dis- 
tributed Understanding in a Multiprogram System, 
Christopher Riesbeck discusses the interactions be- 
tween the various programs developed by generations 
of graduate students at the Yale AI Project. The arti- 
cle is not intelligible in isolation, but anyone who has 
been following the history of MARGIE, SAM, PAM, 
and QUALM will find an interesting discussion of how 
one program leaves notes for another one to elaborate. 
In Five Aspects of a Full-Scale Story Comprehension 
Model Chuck Rieger proposes an integrated system 
for understanding a real story, The Magic Grinder. 
Like Schank and Carbonell, he introduces high-level 
concepts like KISS instead of limiting the system to 
the primitive ACTs of PTRANSing lips. Since some 
of his techniques, such as parsing with word sense 
networks, may have high overhead, his work in testing 
them on a substantial body of text is necessary to 
determine their practicality. 
Various forms of { associative I conceptual I 
cognitive I partitioned I semantic I structured } 
\[dependency linheritance\] {graphs I nets I networks} 
have major differences in notation, terminology, and 
expressive power, but they develop similar network 
solutions for their common problems of representing 
knowledge in processing natural language. Yet many 
areas remain to be developed: 
• None of the systems have a good analog of the 
LISP QUOTE operator. Such a feature would 
enable a system to distinguish between believing a 
proposition and actually changing a model of the 
world. A QUOTE operator is also necessary for 
handling the phrase "dedicated to the proposition 
that..." in Schank and Carbonell's example. 
• Many of the systems have special case inference 
rules for plausible or exact deduction, but ques- 
tions of completeness and consistency are usually 
ignored. Their relationship to various nonstan- 
dard forms of logic (Leblanc 1973) must also be 
investigated. 
• Procedural techniques like Levesque and 
Mylopoulos's will have to be refined and tested 
on more complex examples than factorial. 
American Journal of Computational Linguistics, Volume 6, Number 2, April-June 1980 111 
Book Reviews Discourse Production - A Computer Model of Some Aspects of a Speaker 
• Benign neglect of syntax was once a healthy reac- 
tion to transformational grammar, but the field is 
now mature enough to accommodate more syntac- 
tic sophistication. 
• And for practical systems, naturalness, readabili- 
ty, and ease of use become crucial. 
In summary, this book is an exciting record of work 
in progress. It assumes too much background material 
to be used as an introductory textbook, but it could be 
used in a seminar course on knowledge representation, 
either with supplementary lectures on logic and philos- 
ophy or with programming exercises that show stu- 
dents how to implement such networks. By presenting 
fourteen systems in one volume, it invites detailed 
comparisons that can help knowledge engineers select 
the best features for their future designs. 
John F. Sowa, IBM Systems Research Institute 
Discourse Production - 
A Computer Model of Some 
Aspects of a Speaker 
Anthony Davey 
Edinburgh Univ. Press, Scotland, 1978, 
170 pp., $16.00, ISBN 0-85224-339-1. 
"This book describes a computer program that 
produces English discourse. The program is capable of 
describing in a sequence of English sentences any 
game of noughts and crosses (tic-tac-toe), whether 
given or actually played with the program." (From 
the Preface.) 
The game descriptions have several properties that 
make them non-trivial. The entities of the game are 
referred to by fairly natural English noun phrases, 
making use of anaphora and standard constructions for 
qualification (such as relative clauses), and taking 
advantage of symmetry of the board. In many cases, 
moves are described for their strategic value, and some 
mistakes of the program's opponent are mentioned. 
The following sentence produced by the program illus- 
trates these properties. If you had blocked my line, 
you would have threatened me, but you took the corner 
adjacent to the one which you took first and so I won by 
completing my line. This also illustrates that the pro- 
gram uses connectors like and, but, so, and although, to 
string clauses together in coherent discourse. 
The grammatical theory used by the author is sys- 
temic grammar, essentially the version developed by R. 
A. Hudson in English Complex Sentences, North- 
Holland, 1971. In this theory, a grammatical item 
(such as a clause) is classified by an associated bundle 
of features. Feature-realization rules determine a set 
of functions (like SUBJECT) for constituents. 
Structure-building rules manipulate functions, gather- 
ing them into bundles, one for each constituent. 
Function-realization rules associate feature bundles 
with these function bundles, and the cycle continues, 
down to the word level. On each level there is some 
freedom of choice for the features of an item, but 
there is a network of constraints on co-occurring fea- 
tures called the system-network. 
In Davey's adaptation, semantic representations are 
carried along with the above sort of syntactic repre- 
sentations, and "specialist" procedures are used to 
determine features of grammatical items where there is 
freedom left in the system-networks. The overall con- 
trol, however, is not strictly top-down as in Hudson's 
system described above. Sometimes specialist proce- 
dures actually construct the English text for an item 
before determining all of its features. 
It is unfortunate that the book did not reach pub- 
lished form earlier. The work was done in the period 
1970-1973 as the author's thesis at Edinburgh Univer- 
sity. A chapter devoted to A Review of Previous Sys- 
tems essentially covers only what was published by 
1972. The program does not take account of advances 
made in systemic grammar since around 1972. Some 
of the constructions made by the program seem less 
direct than they should be, and this would be im- 
proved in the light of later developments. 
Michael C. McCord, University of Kentucky 
112 American Journal of Computational Linguistics, Volume 6, Number 2, April-June 1980 

References 
Bar-Hillel, Yehoshua (1960) "The present status of automatic 
translation of languages," in F. L. Alt, ed., Advances in Comput- 
ers, vol. 1, Academic Press, New York, pp. 91-163. 
Ceceato, Silvio (1956) "La grammatiea insegnata alle machine," 
Civilt~ delle Machine, Nos. 1 & 2. 
Ceccato, Silvio (1961) Linguistic Analysis and Programming for 
Mechanical Translation, Gordon and Breach, New York. 
Ceccato, Silvio (1962) "Automatic translation of languages," pres- 
ented at the NATO Summer School, Venice, July 1962; reprint- 
ed in Information Storage and Retrieval, vol. 2, no. 3, pp. 105- 
158, 1964. 
Hays, David G. (1964) "Dependency theory: a formalism and 
some observations," Language, vol. 40, no. 4, pp. 511-525. 
Klein, Sheldon (1965) "Automatic paraphrasing in essay format," 
Mechanical Translation, vol. 8, pp. 68-83. 
Leblanc, Hughes, ed. (1973) Truth, Syntax, and Modality, North- 
Holland Publishing Co., Amsterdam. 
Schank, Roger C., & Lawrence G. Tesler (1969) "A conceptual 
parser for natural language," Proceedings of the First IJCAL pp. 
569-578. 
Tesni~re, Lucien (1959) ElEments de Syntaxe Structurale, Librairie 
C. Klincksieck, Paris, second edition 1965. 
