Book Reviews Conceptual Structures
inferences are generated, how many inferences are Jean-Pierre Corriveau
generated, and what knowledge sources contribute to the Department of Computer Science
generation of inferences. University of Toronto
In their book Structures and Procedures of Implicit Toronto, Ontario
Knowledge, Graesser and Clark, two psychologists, Canada M5S 1A4
attempt to answer these questions by presenting a model
of comprehension that primarily focuses on knowledge-
based inferences (viz. products of what the comprehen- CONCEPTUAL STRUCTURES:INFORMATION PROCESSING
der knows about the world). IN MIND AND MACHINE
The likelihood of a particular inference depends on the (The systems programming series)
content of the inference, together with (a) the text's
John F. Sowa
context, (b) world knowledge structures and inference
engines that are available, (c) the goals of the compre- Reading, MA: Addison-Wesley, 1984, xiv+481 pp.
hender, and (d) the pragmatic context of the communi- ISBN 0-201-14472-7
cation act. Admittedly, this is too much to handle at
once, and the authors do not wait until the last section of John Sowa has written an excellent book. It is beautifully
the book to clearly state the goals and limitations of their written, and presents a clean, precise look at knowledge
work. There is no discussion of syntactic parsing, no representation and its applications. The b o o k combines a
formal theory of meaning, and no pragmatic model. This sweeping historical perspective from the ancients to
is not a book about linguistics but rather about conceptu- current research, with a formal definition of knowledge
al modeling and, specifically, about the generation and representation structures.
usage of knowledge-based inferences during text compre- The first two chapters provide the motivation and set
hension. The authors are quick to point out that formal the tone for the rest of the book. They are a delight to
work in linguistics, logic, and philosophy, as well as AI read. Chapter 1, "Philosophical Basis", shows why
research, ignore "important characteristics .of human psychologists, linguists, philosophers, and computer
cognition". Implicitly, researchers in those 'fields are scientlsts ate all interested in the problem of knowledge
invited to momentarily leave their idealistic vacuums or representation, and how their perspectives on the prob-
their Lisp code in order to refresh their knowledge about lem overlap and differ. One finishes the chapter with an
text comprehension and psychological plausibility. understanding of the historical development in each of
Though the book does not present an exhaustive the areas, and the interdisciplinary nature of the field.
survey of the available research on inferences and text Throughout, Sowa stresses the importance of formal
comprehension, its concise discussion in the first chapter models, as opposed to ad hoc solutions. Chapter 2,
of inference taxonomies and engines and its rich bibli- "Psychological Evidence", surveys numerous psycholog-
ography make it an excellent reference. ical experiments that illustrate the nature of human
Text comprehension is extremely complex, and the language behavior. Modeling this behavior is the essential
authors can only offer a very partial yet quite interesting problem of AI research. The chapter contains several
solution: they propose procedures to model comprehen- wonderful anecdotes - for example, the eidetic-memoried
sion, recall, summarization, and question answering. Shereshevskii who lost his job as a newspaper reporter
These procedures work on generic knowledge structures because he could not abstract from detail.
(represented by conceptual graphs) that they traverse At Chapter 3, "Conceptual Graphs", we enter the
and match in order to generate the inferences that make technical part of the book. This chapter is a fine intro-
the text coherent, as well as other inferences that capture duction to semantic net representation, which he calls
the comprehender's expectations. From a computational conceptual graphs. Conceptual graphs are a canonical
point of view, since there is no implementation of the form of many AI knowledge representation schemes. In
model, the discussion may sometimes appear superficial. the form he follows throughout the book, Sowa first
Also, Graesser and Clark too often claim without any presents a general, understandable discussion of what is
further explanation that their model includes previous to be represented and why, and follows with formal defi-
work. As is usually the case for this domain of research, nitions. He covers all the fundamentals: concepts, gener-
since the stories analyzed must minimize the role of the alization and specialization, types and tokens,
components left out of the model, the reader is aggregation and individualization. By the end of the
confronted with truly artificial and insipid texts. chapter we have a good intuitive understanding of the
Finally, be forewarned! The authors present a gener- representation, and a sound formal basis.
ous amount of statistics obtained from numerous exper- Chapter 4, "Reasoning and Computation", proceeds
iments; a great deal of time is spent analyzing the data naturally from Chapter 3. We see larger organizations of
and defending the methodology. The reader may often memory structures, and how they are used in reasoning
want to skip to the end of chapters, where good summa- processes. He shows how logic can be represented in
ries of the results and conclusions are provided. graphs and how deduction can be performed on them.
218 ComputationalLinguistics, Volume 12, Number3, July-September 1986
Book Reviews Conceptual Structures
Chapter 5 is on "Language". Again, we have an excel- As a textbook, Conceptual Structures has both positive
lent, broad summary: child language abilities, strata of and negative aspects. On the positive side, it gives an
language, case grammar, and generation. There is a nice historical perspective of many fields, presents a general,
presentation of syntactic analysis. However, I would like formal definition of knowledge representation and its
to see a deeper discussion of ATNs; here he just mentions uses, and contains good exercises. An additional bonus is
the phrase. The sections on context, and integrating the conceptual catalog given in Appendix B. This is
syntax and semantics are particularly well done. exactly what students always ask for: a collection of
Chapter 6, "Knowledge Engineering", is a bit of a structures that can be used by programs. On the negative
grab bag. On expert systems, he mercifully avoids the side, specific AI systems are nowhere discussed in detail.
current hype by taking a "Just the facts, Ma'am" atti- It would be necessary for the instructor to provide
tude. The other sections on natural language systems, supplementary material and original sources.
database semantics and inference, knowledge acquisition, I agree with Clancy (1985) that "Every AI and Cogni-
and learning are all informative, succinct descriptions of tive Science researcher should study the conceptual graph
basic problems and the current state of the art. notation and understand its foundation in logic, database,
Chapter 7, "Limits of Conceptualization", ends the and knowledge representation research."
book with a description of the wonderful aspects of
human behavior that computers can't do at all. It is great Sharon Salveter
fun to read. Department of Computer Science
The potential audience for this book is extensive. Boston University
Because the chapters are organized with introductory 111 Cummington Street
material at the beginning of each section and formal Boston, MA 02215
descriptions later, readers may skip sections without Reference
losing the thread of the book. A newcomer to AI can
Clancy, W.J. 1985 Review of Conceptual Structures: Information Proc-
read the first parts of each section, serious students can essing in Mind and Machine by J.F. Sowa. Artificial Intelligence
read it all. 27(1): 113-124.
Computational Linguistics, Volume 12, Number 3, July-September 1986 219
