Jeng, 'Knowledge Acquisition: Eliciting Human Expertise for Building Expert Systems', LITA Newsletter v14n04 URL = http://hegel.lib.ncsu.edu/stacks/serials/lita/lita-v14n04-jeng-knowledge [v14n4.knowledg litanews] ------------------- Knowledge Acquisition: Eliciting Human Expertise for Building Expert Systems Judy Jeng Knowledge acquisition, the heart of expert system development, was the theme of this program, sponsored by LITA's Artificial Intelligence/Ex- pert System Interest Group. The program featured three speakers and one demonstration. Ralph Alberico Alberico (University of Texas at Austin) opened the program by discuss- ing the nature of expertise, the central problems of knowledge acquisi- tion and the two major approaches to knowledge acquisition. He then re- viewed a number of specific knowledge-acquisition techniques. According to Alberico, expertise is not the same as textbook knowl- edge. It is not as neat, not as well categorized. Expertise is mostly unrecorded, mostly in the heads of human beings. Expertise is diffi- cult to articulate. Expertise is compiled in the minds of experts. Ex- perts are hard-wired to their disciplines. Their expertise is burned into their minds and therefore is difficult to decompile. Expertise is based on experience. Experts don't have to think about what they are doing, often can't explain what they are doing or why, and know their limits. They are able to understand when they are approaching the boundary of their knowledge and make referrals. They are also able to use their body of knowledge to deal with unique and unfamiliar circum- stances. Sources of knowledge include human experts themselves, observations of human experts, the literature, databases and written procedures. Types of knowledge are: cognitive, procedural, declarative, common sense, creativity, social knowledge, meta-knowledge (knowing what one knows) and consensus knowledge. An ideal automated knowledge acquisi- tion could be charted as in Figure 1 (omitted in Internet version). Two major approaches to knowledge acquisition are the symbolic ap- proach and the connectionist approach. The symbolic approach acquires and represents knowledge by manipulating symbols. It focuses on spe- cific knowledge. It is knowledge-based, structural, psychological, de- ductive and linguistic. It is labor intensive and relies heavily on in- terviews. It is a "say-how" approach. On the other hand, the connectionist approach is pattern-based, statistical, mathematical, in- ductive and non-linguistic. It is computationally intensive and typi- cally based on observations. Alberico reviewed a number of specific knowledge-acquisition tech- niques. They are structural interviews, protocol analysis, repertory grid techniques, machine induction, ID3 algorithm and "mining" data- bases. Knowledge acquisition is widely acknowledged as the great bot- tleneck in expert system design. It is a problem that has resisted easy solutions. Finally, Alberico speculated that the future direction of knowledge acquisition will be more connection, more hybrid approaches, more con- sensus knowledge bases, more knowledge from the literature, more knowl- edge extraction from databases and more agents to acquire knowledge. Tschera Connell Connell (Kent State) discussed her application of a combination of knowledge-acquisition techniques to analyze the online search proc- esses that naive searchers employ when using automated information re- trieval systems. The selected methods of eliciting knowledge include task-oriented techniques and interview-oriented techniques. Lynn Branche Brown Brown (Penn State University) demonstrated her prototype system that advises whether a particular title will come via approval plans. Deter- mining factors are the publisher, the LC class, cost (less than $200), year of publication and place of original publication. Comments This program attracted some 50 people who were interested in knowl- edge-acquisition techniques and their application in library and infor- mation science. Although some parts of the presentation were technical and might be dry for some audiences, the program was well organized. Speakers were knowledgeable. The demonstration of an approval plan ex- pert system especially added a flavor to this program.--Judy Jeng is Head of Collection Management/Technical Services & Media in the John Cotton Data Library, Rutgers--Newark.