|Volume 1: No. 03|
The following were among the tools and prototypes scheduled for demo at RIAO 91 ("Intelligent Text and Image Processing") in Barcelona early this month [Ed Fox (email@example.com), NL-KR]:
- ALCESTE text-analysis system (IMAGE, France) - PRIAM real-time news bulletin analysis (SY-MEDIA, France) - AMI French/German message sorting (Thomson RCC and CORA, Fr) - STO3/STO4 Speech Recognizer System (G. Bekery Acoustical Research Laboratory, Hungary) - SGML/Search retrieval system (Berger Levrault--Advanced Information System, France) - ZEN X-Windows hypertext system (BULL, France) - GUIDE, IDEX (OWL, Great Britain) - TOPIC graph-based document retrieval system (Verity, USA) - MULTIMEDIA INFORMATION SYSTEM (IN TECS, Italy) - TAKE 5 multilingual pseudo-NL retrieval system (EDIAT, Fr) - SPIRIT information retrieval system (SYSTEX, France) - DOXIS structured/full text system (ERGOSUM, France) - STATUS/E structured/full text system (HARWELL COMPUTER POWER LTD, Great Britain) - PSIDOC software tools for documentary databases (JOUVE, Fr) - MOVIE retrieval with interactive decision making (COSM, Fr) - ILIADA retrieval and library management with full-text and hypertext support (Software AG, Spain) - MULTIMEDIA SYSTEM for teaching biological sciences (CTU Univ. of Milan, Italy) - METAL machine-translation system for several languages (Siemens, Germany, Spain) - TEXIRIS 2 PLUS Omnifont OCR with 12-language dictionary (Image Recognition Integrated Systems IRIS, Belgium) - READSTAR OCR on a TRANSPUTER card (Inovatic, France) - TEXTPERT OCR for MacIntosh and Windows (CTA, Spain) - Intelligent Information access to 200 data banks in Europe and the U.S. (INFOTAP, Luxemburg) - EPOQUE access to European Patent Office Databases (European Patent Office, CEE Holland) - IMAGEDB image/descriptor database management (CSI, Spain) - AUTOMABB PC multimedia authoring system (MABB Sistemas Interactivos, Spain) - INFODOC document storage, image processing, and retrieval (INFODOC SA, Spain) - CLARITY text/image archiving system (Micronet, Spain) - HYPARCHIV Windows hypertext (ACS Systembereitung, Germany)
Review -- SKE knowledge-engineering methodology.
Karen N. Gardener of the Bechtel AI Institute gave a talk on Structured Knowledge Engineering at a recent AIA meeting. SKE is a software-development methodology with enough structure and documentation to be acceptable to MIS managers. Bechtel AI Institute is teaching short courses based on the work at BOLESIAN, B.V. in The Netherlands, a firm that will soon grow from 35 to 300 knowledge engineers. [AIA is the Artificial Intelligence Association, a monthly seminar group that meets in Palo Alto. The contact is Leor Jacob, (408) 734-5760. Bechtel is at 50 Beale St., P.O. Box 193965, San Francisco, CA 94119-3965.] SKE diagrams are similar to Entity-Relationship diagrams, but with "inference boxes" and information-flow arrows replacing relationships. These diagrams represent the implementation of "tasks" in terms of "inference modules," two of the four knowledge layers that are iteratively refined during algorithm design. A top layer addresses control issues; a bottom layer maps between SKE terminology and hierarchical domain-knowledge representations. Once the conceptual design is complete, implementation can take place in any computer language or shell. (Bechtel is currently evaluating the strengths and weaknesses of different expert-system tools for implementing each type of knowledge representation and inference process.) A strength of this approach is the catalog of two dozen or so standard inference modules (categorization, diagnosis, selection, refinement, etc.) Anyone learning to think in these terms would indeed be helped in high-level algorithm design. It's a sure cure for programmer's block. A weakness is that the repertoire has not yet expanded into continuous domains such as signal processing. All of the examples I saw were suitable for discrete manipulations (financial applications, object-oriented programming, logical inferences). That's fine for the bulk of all programming, but I saw nothing that would support the design of neural networks, blackboards, or other signal-processing techniques. Karen also seemed to find clustering and multidimensional mapping techniques difficult to think about in this framework, except as black-box "heuristic matches." Low-level knowledge relationships were quite restricted, consisting only of attribute, sequence, is-a, and caused-by. Representing different types of is-a relationships could be quite difficult. (Clyde is an elephant. Elephant is a species. Therefore Clyde is a species.) My conclusion is that SKE will indeed be an important methodology, offering a bridge between MIS and AI people. I'm no authority on systems analysis and functional decomposition, but SKE appears to be an attractive alternative. It is certainly preferable to software design approaches that are limited to a particular expert-system shell. I do regret, though, that Karen was so negative about AI's rapid-prototyping approach. Prototypes are very useful exploration and marketing tools, and I would hate to see them banned in software development contracts.