Findler describes his work on the development of the System for Heuristic Retrieval of Information and Facts (SHRIF). Very much in the AI mold, the author starts by noting the cognitive proficiency of humans in such areas as assimilating, categorizing, evaluating, and searching for information. He believes SHRIF is a good start toward approximating these human proficiencies in the application of information retrieval as well as providing the basis for information systems in such applications as diagnosis, troubleshooting, classification, and causal reasoning. While some interesting techniques are associated with SHRIF--such as an interactive mode for adding nodes to the knowledge representation (KR) scheme and clustering algorithms for the KR nodes--the demonstrated results of its use to date indicate no more than a modest start toward achieving its objectives.

After the introduction (chapter 1), a chapter on the “Concepts of Information Retrieval” provides slightly more than a bare definition of such terms as document and fact retrieval, precision and recall, database management systems, information storage and retrieval systems, and question-answering systems. Chapters 3 through 5 provide cursory reviews of example computer systems in the fields of information retrieval, natural language, and medical information. Readers who are not already acquainted with these areas will not learn much on which to base their understanding of what follows. In particular, essentially nothing done in the last ten years is mentioned.

Chapter 6 describes eight clustering algorithms in some detail. This chapter provides an interesting discussion, at a reasonable level of detail for a non-expert, of several clustering techniques. Nine references for further reading are given.

Chapters 7 through 12 and the appendices (the remainder of the monograph other than a short conclusion and the references) give a detailed description of the SHRIF system, including its design, several components, and examples of its use. The SHRIF knowledge representation language is based on ten descriptor (relationship) types: attribute-value (for example, for the concept node “back pain,” the attribute “medication” has the value “ampicillin”); synonym; superset; subset; commonality (nodes sharing common terms or properties); exceptionality; cause and effect; and process affecting, derived from, or regenerating some entity. Defining a few additional attribute-value relationships such as symptoms, causes, and treatment, the author shows how some presumably simple common facts about two diseases (acute pyelonephritis and asthma) can be represented in this language. An important feature of the SHRIF system is the association of the concept nodes into affinity groups called clusters and the further association of clusters into planes. Based on the author’s analysis of a number of associative techniques, the SHRIF system allows KR builders to choose from several association algorithms and, beyond that, choose levels of affinity by which the strength of association and the size of clusters can be specified. The query is basically a relational triple in which a relation connects an object and a value, where two of these three are specified and the third is the unknown to be found. Thus, for the (English form) query “What are the symptoms of pneumonia?” the value of the relation symptoms is sought for the object pneumonia.

This monograph is well laid out, typographically clean and error-free (as far as I could tell), and easy to read. It nicely explains the SHRIF system through verbal and more formal descriptions as well as KR and usage examples. Because the monograph is a clean, relatively easy read and has a host of references (about 150), it might be a nice way to introduce a student to one or more of the fields discussed and encourage him or her to look beyond the rather superficial levels found in the monograph itself. (Perhaps the best review of early medical information systems, by Clancey and Shortliffe [1], is not in the reference list, however.) Serious researchers in any of these fields may have several reasons for reading this text. The discussion of the various clustering techniques could be useful. The KR language itself might be studied by those embarking on a similar enterprise to check that all the relational types in SHRIF are covered. The SHRIF interface, which allows a KR builder to enter concepts in an online mode, with a certain amount of natural language analysis built in to assist this process, appears to be a nice attempt in this direction.