Site: NTT Yokosuka Laboratories
1-2356 Take Yokosuka-shi
Kanagawa-ken 238-03, Japan

Date Visited: March 23, 1992

JTEC Attendees:

Friedland
Johnson
Shrobe
Chien

Hosts:

Dr. Tsukasa Kawaoka

Executive Mgr.,
Knowledge Systems Laboratory
NTT Network Information Systems Laboratories

Mr. Daiji Nanba

Senior Research Engineer,
Supervisor

Mr. Toshiyuki Iida

Senior Research Engineer,
Supervisor

Mr. Fumio Hattori

Senior Research Engineer,
Supervisor

Dr. Satoru Ikehara

Senior Research Engineer,
Supervisor

Mr. Francis Bond

ORGANIZATIONAL DATA

NTT's Knowledge Systems Laboratory has about 100 researchers -- 80 scientists and 20 engineers (perhaps "technicians" was meant rather than engineers). There are about 200 other personnel in the company doing actual application building.

PRESENTATIONS

The meeting consisted of two parts. The first part was a question and answer session on NTT operations, the company's expert system development tool "KBMS," and answers to the questionnaire. The second part consisted of a series of demonstrations of expert systems in the areas of private network design, intelligent tutoring, machine translation and a data entry system for service orders.

SUMMARY OF QUESTIONNAIRE

NTT has currently 36 expert systems under development. Half of these systems perform diagnosis on various components of a communications system. The company's most successful ES performs failure diagnosis and support of crossbar switching equipment. A task that typically takes four hours has been reduced to five minutes using the system. However, the main motivation for developing the system is that they are going to phase out the rest of the crossbars and no longer want to train people on it. With the expert system they don't need experts.

There is also research in three major areas: common sense AI (but note a quite specific definition of this as limited to reasoning quantitatively about ambiguous concepts and modifiers), machine translation, and VLSI design (high level description going to synthesis methods).

The work on common sense AI focuses on making quantitative judgments in very large KBs -- e.g., how long numerically a "long" river is. This also includes judging ambiguity in words, e.g., a long vacation vs. a long pencil. The machine translation work is based on belief in the great importance of specialized knowledge. Currently the system has 15,000 sentence structure descriptions.

NTT has developed a company-proprietary process for ES development.

OBSERVATIONS

Our overall impression was one of considerable vitality and high morale. NTT did a meticulous job in filling out our questionnaire for both applications of KB systems and more basic KB research. Major questions the answers which remain ambiguous are precise quantitative payoffs of currently fielded systems and whether the research work described adequately explains the activities of a 100 person laboratory.

Diagnosis is the biggest area, by far, of application development. The biggest ES "win" for NTT has been a crossbar diagnosis system. This is fully automatic and integrated with trouble-reporting systems. Dr. Kawaoka stated that "the battle for expert systems has been won at NTT."

The JTEC team saw demos of four systems: design of local switching networks, verification of service orders, machine translation, and an intelligent tutoring system. This latter work seemed fairly simple and not at all state-of-the-art. The switching network design system used straightforward heuristic methods (as opposed to a combination with linear programming methods), the service order verification work was nicely tied to an optical character recognition system, the machine translation work was impressive.

A 100-person AI lab is very large by any standards, and represents a huge corporate commitment. Moreover, these are not the same people who build the applications. The technology was generally mid-1980s vintage.

None of the applications are simple "advisors."

NTT's perspective appears to be that all knowledge is heuristic and is mined from experts, rather than having a more formal problem formulation (such as model-based reasoning) or a more automated knowledge acquisition technique (such as case-based reasoning). It was surprising that there is no ongoing research in model-based diagnosis, given the corporate emphasis on diagnosis.

NTT's analysis of the benefits did not seem to rely heavily on cost/benefit justification but rather upon incremental quality/service improvements. In fact this is a possible hypothesis for the JTEC study: a high percentage of Japanese expert system initiatives reach the operational stage because the Japanese value incremental quality and service improvements which may not have the easily quantifiable cost/benefits expected by U.S. businesses. Expert system technology seems well suited to solving problems which produce these "soft" benefits.

Computer-based Instruction: Individual frames are represented with crude computer graphics or with analog audio or video. No DVI is employed. The application relies upon manual indexing to establish goals and subgoals and to map the frames of training material to the subgoals. No case-based indexing is used. It was not clear to us how the user model interacts with the selection of subgoals. Forms oriented data entry is the norm. This is not state of the art, namely graphical course maps which allow direct manipulation and which rely upon an underlying object-oriented design, which the NTT application apparently does not employ. In general this application lacks the features which will capture the imagination and provide motivation to students. This may not be very important in Japan, given the work hard/study hard ethic. It would be a fatal flaw in the U.S. The JTEC team did not see any indication that this application has been installed.

Service Order Entry system: This application inputs manually encoded information, uses a scanner and optical character recognition (OCR) to encode the information, and supports manual verification through on-screen comparison of the image with the encoded information. Next the order is checked for consistency. Finally, the system sends a transaction in some form to the service order transaction processing systems. This is a basic application which is installed in a large number of locations. It operates on PCs with a scanner and OCR box attached.

Network design system: This is a synthesis problem, but was primarily solved through heuristic methods.

Crossbar switch diagnostic application: NTT has developed its own Knowledge-base Management System (KBMS). They call it a second generation KBMS. KBMS is marketed separately. The system is comparable to KEE.

NTT is developing a Japanese to English translation system which is impressive. It is not clear that the particular approach will scale up, however.

The expert systems we saw were pretty typical of what you would see anywhere. They seemed to be at the same state of the art as in the U.S. Overall, the quality of the engineering work was very good.

The number our hosts gave for percentage of applications that progress from prototype to deployment is a staggering 95 percent! They staff the average project more heavily than we typically do in the U.S. -- about 10 people seemed average. Finally, they said that quality control was a primary source of system suggestions; again quite different from the U.S.


Published: May 1993; WTEC Hyper-Librarian