SIGART’s special refereed issue of its quarterly newsletter,devoted entirely to knowledge acquisition, reflects the state of thefield of knowledge acquisition. It is therefore appropriate to make someprefatory remarks about the field before assessing the issue’scontribution to it.

Although the design and development of expert systems is arelatively new endeavor, commercial success and scientific achievementhave propelled the field forward at an unusual speed, and expert systemsare often thought of as the “bread and butter” ofAI. One would think, therefore, that knowledgeacquisition, methodologies for acquiring the knowledge necessary toproduce the knowledge base on which an expert system depends, would besimilarly advanced. However, this is simply untrue. Although developingstrategies for knowledge acquisition is a task logically antecedent tothose directly concerned with the design and development of expertsystems, it is only now that the field is being recognized.

Among academic journals of note, only theInternational Journal of Man-MachineStudies has covered the field for any length of time.Three new journals--KnowledgeAcquisition in Great Britain, Dataand Knowledge Engineering from Elsevier,  and theIEEE Transactions of Knowledge and DataEngineering--have just opened their pages, andACM has just recognized the emerging field with this special SIGARTissue. Since an almost 15-year time lag exists between the emergence ofthe field as a science and the appearance of these publications, itdeserves some explanation. Computer scientists are accustomed toproducing specialized software and tools to develop software; in thatsense, expert systems presented only challenges that were amenable toextensions of available techniques. Knowledge acquisition, on the otherhand, has almost exclusively been the extraction of knowledge from adomain specialist by a knowledge engineer, a subfield called“knowledge elicitation,” a subject which, in itself, hasless to do with computer science than with psychology and education.Computer scientists have not only not been trained to think in terms ofalgorithms for such interpersonal tasks; they have not even been trainedto notice the task.

Nevertheless, it soon became evident that the “knowledgeacquisition bottleneck” was overwhelming entire expert systemsprojects, and that forced attention to a task not normally conceived ofas scientific. In this newsletter, the contributors agree on thecentrality of and difficulties inherent in knowledge elicitation. Boose[1] has called knowledge acquisition not just a “majorproblem” but also the “central task” in the design ofan expert system. M. LaFrance, in his contribution, “The qualityof expertise: understanding the differences between experts andnovices” (pp. 6–14), quotes Duda and Shortliffe, bothpioneers in the theory and development of expert systems, to the effectthat “expertise elicitation [is] one of the most complex andarduous tasks encountered in the construction of an expertsystem.” N. Shadbolt and M. Burton, in “The empirical studyof knowledge elicitation techniques” (pp.15–18), describethe task as “vexing,” while N. Lavrac, in “Methods forknowledge acquisition and refinement in second generation expertsystems” (pp. 63–69), finds it a “demanding mentalprocess,” and B. W. Crandall, in “A comparative study ofthink-aloud and critical decision knowledge elicitation methods”(pp. 144–146), identifies it as a “major hurdle.” R.Hoffman, in “A brief survey of methods for extracting theknowledge of experts” (pp. 19–27), goes so far as to callknowledge elicitation “exhausting.” It is therefore easy toconclude, as do P. F. Micchiche and J. S. Lancaster, in“Applications of neurolinguistic techniques to knowledgeacquisition” (pp. 28–33), that “interviewing, despiteone’s initial impressions, is a challenging task.”

Finally, the problem has been recognized. It has not, however, beensolved and the newsletter’s contributions to the literature on knowledgeelicitation are not any more helpful than the rest of the literature inthis brand-new field: that is, not atall. Technological aids, observations duringelicitation, transcripts, grids--all of these fail to appreciablylessen the knowledge acquisition bottleneck. None offers us an algorithmfor interviewing which is effective (even generally) and definite (evenup to a point). In reading the individual contributions, I was mostreminded of the pre-Socratic philosophers’ scientific theories: amixture of speculation, ad hoc methods and techniques, and an occasionalinsight. As a human factors field, knowledge elicitation isstill at square one.

Perhaps not surprisingly, the field has advanced only throughautomation techniques, rather than person-to-person interviewing. Thesetechniques provide for the computer to learn from the domain specialistwith relatively little or no help from a knowledge engineer. Computerinduction is, of course, a well-established branch of AI, and itsapplication to building a knowledge base is not a radical departure frommore traditional computer projects. For examples of successful work seeMichie [2], Quinlan et al. [3], and Burzesi [4]. Even here, however, thenewsletter fails us. Its contributors offer little beyond what hasalready been published and in some cases, such as Lavrac, even genuinelyengineered methodologies presented well elsewhere are presented here inalmost abstracted form. I cannot see any knowledge engineer reading oneof these short pieces and coming away with a specificmethodology--even one that is not properly described asalgorithmic--of help in an actual project. Nor will a reader comeaway with much insight. I can only recommend this collection to thosenew to the field who wi sh to see where the field stands--pretty ornot--and they will come away disheartened by being required to plodthrough the jargon that is substituted for methodology, but pleased thatthey have found a field in which there is clearly room to makesubstantial advances.

ACM implicitly recognizes that progress has been made only withautomation, not interviewing, since the Classification System includes“knowledge acquisition” only under I.2.6(“Learning”), which deals with machine induction. As theinterview method goes beyond its initial basis, it will be appropriateto include nodes in H, I.2.1, K.6.m, H.1.2, and H.3. For now, however,the judgment implicit in the Classification System and explicit in thisreview of the newsletter seems to reflect the state of the art quiteaccurately.