A Remedy for Risk
Rooting Out Character Errors
Microdrive Breaks Through the Size Barrier
IBM and Ford Get Down to First Principles
Remedy for Risk
Amidst the recent economic turmoil in Japan, the country's investors and analysts have a tool that can help take some of the risk out of doing business. IBM's new Default Meter software predicts the probability that a small or medium-sized Japanese company will go bankrupt, and displays the result in simple-to-read fashion.
The program fills a previously vacant niche, according to Akira Tajima, a researcher whose team at IBM's Tokyo Research Laboratory (TRL) developed Default Meter. Programs such as JP Morgan's CreditMetrics™ or Standard and Poor's rating systems are designed to estimate the financial health of large, listed corporations, not smaller, unlisted ones, Tajima says. "And while there are many existing technologies to assess portfolio management, or derivatives, options and other risks, none of these really address the problem of credit risks."
Nor is a company's stock price a dependable guide for analysts, Tajima contends. "Computing the reliability of a Japanese company is made exceedingly difficult as stocks are not properly valued in Japan, mainly because disclosure is not sufficient. As a consequence, many people don't believe the price. It's a problem."
IN THE GREEN
The raw material for Default Meter is company balance-sheet data supplied by the Teikoku Data Bank. Among other things, the software uses this data to identify rules for discriminating between safe and risky companies. It considers factors such as equity
ratio versus debt interest rates. The software can then produce company ratings for one, two and three years. The company's outlook is graphically pinpointed on a green-and-yellow graph. If the company falls within the green, it's safe to assume it is creditworthy, but if it falls in the yellow, investors or lenders are advised to watch out.
The program, which runs on
Windows® 95 or NT® platforms, has aroused considerable interest from banks and financial companies since its introduction in April 1998. This is partly because of its simplicity and partly because of fortuitous timing, says TRL's Yasuhiko Morimoto. "In the same week that the system was first publicly announced, a major securities company went bankrupt," he says. "It was bad for them but resulted in a lot of coverage. We have opened a
window for Japanese companies that wasn't there before."
-- Paul Kallender
Rooting Out Character Errors
For most of us, the joy of learning to write was partly undone by the duty of learning how to spell. Today, computer spellcheckers offer a safety net that even the most diligent dictionary users find invaluable. Not only is a spellchecker faster than looking up a word, it also catches all those words you thought you knew how to spell. Invoked interactively or when one has finished a document, it inspires a sense of virtuous conscientiousness.
A somewhat different problem of checking a document arises in character-based languages such as Chinese. Typically, one does not type characters stroke by stroke, so there is little chance of miswriting one. But other types of errors can creep in, and scientists at IBM's China Research Laboratory in Beijing have developed advanced techniques to detect and correct them -- techniques available in a software package called CEC (Chinese Error Correction) 1.1 for Lotus Word Pro®.
HOW ERRORS OCCUR
Because many characters sound or look alike, errors in Chinese text are numerous. Ironically, entering text on a computer introduces more errors than occur in handwritten documents. It happens because a common way of inputting text is to use a phonetic transliteration of the desired characters. The result is a list of all the characters with the same sound, and the user might, through carelessness or ignorance, choose the wrong one.
In general, Chinese characters possess more meaning, or semantic content, than English letters but less than English words. It is therefore almost impossible to detect errors in a Chinese sentence by examining individual characters. To complicate matters, Chinese words are not separated by spaces.
The first step in error-checking Chinese is to segment the text into the most likely set of words. "Segmentation, which determines the word boundaries in a sentence, is carried out with the help of a Chinese word dictionary that contains information about the frequency of each word in a large corpus of texts," says Qiu Zhaoming, a member of the team that developed CEC.
Various possible segmentations of a sentence are scored according to how common the resulting words are. The segmentation with the highest score is then checked for so-called lone-character pairs. Words represented by a single character are relatively rare in Chinese, and the incidence of two such characters together is even lower. Since the two characters together do not constitute a word in the CEC dictionary (otherwise they would be segmented as such), they are checked independently, using a special "bigram table," which indicates the likelihood that any two lone characters will appear beside each other. Lone-character pairs below a certain cut-off value are marked as possible errors.
But these techniques still can't catch another common type of error, the misremembered idiom, notes Guo Zhili, a member of the CEC team. The software therefore includes a list of common expressions, against which variants in a text can be checked and corrected.
For simplified Chinese, the form of written language used in the People's Republic since the late 1950s, CEC 1.1 can check 15,000 to 20,000 Chinese characters per second, the equivalent of about 10,000 English words. That makes it about 10 times faster than one leading Chinese error-check product. Its detection rate is also 20 percent better. While another popular error-check product is twice as fast as CEC 1.1, its false-alarm rate is almost twice as high and its detection rate 30 percent lower.
Overall, the market for a tool like CEC 1.1 is vast. China's 1,000-plus newspapers,
nearly 8,000 magazines and 120,000-title-a-year book industry could all benefit.
The software may not put proofreaders out of work, but it will make their job --
and the task of every other computer user -- considerably easier.
-- Xu Fang
Microdrive Breaks Through the Size Barrier
IBM HAS UNVEILED the world's smallest hard disk drive: it's no bigger than a large coin. Weighing less than an AA battery, it holds 340 megabytes -- 200 times more data
than a floppy. The microdrive is ideal for use in pervasive computing devices, such as digital cameras, handheld PCs and cellular phones, and will allow data to be shared among
the different devices.
"We realized that shrinking down our disk-drive technologies could greatly enhance the capabilities of portable electronics," says Tim Reiley, the Almaden Research Center
scientist who initiated the effort. "By working closely with Storage Systems Division development labs in San Jose and Japan and with the Compact Flash Association, we
created an innovative design we believe will be both pioneering and successful."
Unlike other tiny storage devices, such as flash memories, the microdrive uses hard disk technology, resulting in much lower per-megabyte costs and five times the storage
capacity. Look for the product in mid-1999.
FYI:
http://www.ibm.com/storage/microdrive
IBM and Ford Get Down to First Principles
Accurate chemical simulations receive their first industrial workout
A recent computer simulation performed jointly by IBM and Ford yielded new insights into an important chemical process. But even more significant than the results
themselves is the way they were obtained. In modeling the interaction between water and alumina (aluminum oxide), scientists at IBM's Zurich Research Laboratory (ZRL)
and Ford Research Laboratory in Dearborn, Michigan, used a technique that calculates reactions from first principles. That is, the technique accurately represents the electronic
properties of individual atoms and applies the laws of quantum mechanics to predict how they will behave. In contrast, conventional molecular dynamics simulations use
empirically derived models that can only approximate the forces between the atoms.
The technique used in the latest work was developed in the late 1980s by Roberto Car, now at the University of Geneva, and Michele Parrinello, now at the Max Planck
Institute in Stuttgart, and formerly of ZRL. Researchers at the Zurich lab have become experts in applying the Car-Parrinello method to a variety of problems, including the
properties of silicon and organic light-emitting materials and the chemistry of fullerenes. But the Ford/IBM simulation, reported in Science (October 9, 1998), represents the
first time the method has been applied to industrially useful chemical reactions. The research was carried out by Kenneth C. Hass and William F. Schneider of Ford and
Alessandro Curioni and Wanda Andreoni of IBM.
A WETTER IDEA
The current work arose when Ford and IBM agreed to collaborate on research using
IBM's RS/6000® SP™ high-end parallel computer.
"We asked the Ford researchers to
propose some topics for joint work, and they suggested that we study alumina (aluminum oxide) surfaces and their interaction with water," explains Andreoni, project leader at
ZRL. Alumina is interesting to the auto industry both because it is important in the adhesive bonding of
aluminum alloys and because it is applied as a support for catalysts in emission control. The reaction with water is important for several reasons, including that it is significant
in the chemical reactions that lead
to corrosion.
These were the most complex simulations yet attempted with the Car-Parrinello method, since to get realistic results, 135 aluminum and oxygen atoms, as well as up to 10
water molecules, with a total of about 1,000 electrons had to be modeled. The simulation modeled how the molecules behave over a trillionth of a second, dividing that time up
into 10,000 separate steps.
The results include several key findings. One is the answer to the questions of whether water molecules break apart as they bond to the alumina surface. The simulation
showed that they do, very readily. A surprising conclusion is that a wet alumina surface acts quite differently from one that is merely moistened by atmospheric humidity. "We
also showed that, even on a perfect surface with no defects, reaction rates are very high," Andreoni points out -- a result that would have been almost impossible to obtain
experimentally because of the difficulty in producing such perfect surfaces.
-- Eric Lerner
