Browse Books

Reviewer: Christopher K. Riesbeck

This is a very odd—but good—book. Its scope is almost as broad as that of any general text on Artificial Intelligence (AI), but it is not an overview of issues and research in AI. Instead, the author applies a particular formal notation, conceptual graphs, to a wide range of problems. These conceptual graphs are part of the large family of graphical notations used in AI, primarily for knowledge representation and grammatical structures. Unlike most authors, Sowa is most concerned with demonstrating the logical adequacy of his graphs, extending and defining them to include not only first-order logic, but modal logics as well. He even uses them to represent computation a la dataflow. This is the strength of the book. Even if Sowa fails to convince you that conceptual graphs are preferable to linear notations, he definitely makes them a serious contender. At the same time, he raises most of the important representational problems that have to be handled by any other notation. The book begins and ends with its weakest material. Chapters 1 and 2 cover the philosophical and psychological motivational background. This coverage, despite its brevity, raises more issues than the rest of the book can hope to deal with. The core of the book rests in Chapters 3, 4, and 5. Chapter 3 defines most of the essential features of conceptual graphs. Two notations are given, one using two-dimensional networks, and the other using a bracketed linear notation. For example, the conceptual graph (in linear notation) “[CAT] ? 1-2-(STAT) ? 1-2-[SIT] ? 1-2-(LOC) ?:- 9T1-2-[MAT]” connects the concepts “[CAT],” “[SIT],” and “[MAT]” with the relationships “(STAT)” and “(LOC)” to represent the description, “a cat that is in the state of sitting on a mat.” Conceptual graphs are used to represent propositions. Concepts and conceptual relationships fit into type hierarchies, e.g., “[CAT]” has the type “[ANIMAL]” as a higher type, so this graph could be generalized into (among others) “[ANIMAL] ? 1-2-(STAT) ? 1-2-[SIT] ? 1-2-(LOC) :- 5E 1-2-[PHYSOBJ].” Topics covered include individual names, canonical graphs, generalization, specialization, abstraction, and definitions. Much of what is presented is standard semantic network material, but it is fleshed out and formalized with great care and clarity. Chapter 4 continues this development to include schemata, quantification, deduction, model semantics, modal logic, and dataflow. Chapter 5 links this material to language, including case frames, augmented phrase structure grammar, generation, and parsing. Chapter 5 ends with an informal discussion of some of the hardest issues in natural language processing, i.e., context, conversational implication, and metaphor. From this point on, the book becomes much less technical. Chapter 6 is a reasonable survey of issues in applied AI, including expert systems, natural language interfaces, conceptually based systems analysis, and knowledge acquisition. The final chapter is a surprisingly strong criticism of what has gone before, stressing the significant limitations of discrete conceptual representation systems. Appendix A gives mathematical underpinnings of conceptual graph theory. The coverage is much too brief to learn from, but is probably adequate as a minirefresher course. Appendix B is a catalog of the concepts and conceptual relationships used in the text. Appendix B isolates what I think is a major weakness of the book. Appendix B begins by noting the difference between a theory of form and a theory of content. The book has been concerned with the former, but evaluating the theory's worth rests on just how well the forms work when applied to particular domains. The catalog of concepts and relationships listed is much too incomplete and scattered to allow any evaluation based on content. It covers no coherent subset of meaning in any way. For example, it presents a few communication concepts, such as “[COMMAND],” “[COMMUNICATE],” “[MESSAGE],” “[TEACH],” and “[ORDER],” but does not relate them to others presented in the text, such as “[WRITE]” and “[ASK].” To some of us in AI, what is of most interest is working out what the relationships are between concepts of communication, knowledge, action, and so on. Sowa understands this, and points it out in several places; but by following linguistic tradition and using a wide variety of unrelated examples to illustrate different points, he has missed an opportunity to make a contribution to any particular theory of content. Still and all, the breadth of material covered, and (in the central chapters) the depth to which the material is treated, is amazing. There are many surveys of AI in print, and a handful of volumes on programming techniques, but very few theoretically rich textbooks. Sowa has managed to produce one. I recommend it to all students interested in knowledge representation.