IT'S been said that a badly written book is a blunder, but a bad translation is a crime. Consider just one gaffe, this one more comic than tragic, that occurred when an Arab once translated a British news dispatch into his own tongue:
''Four coaches of the train were telescoped into one another'' came out, ''Four coaches of the train were flung so far that it required a telescope to see them.''
Since the early 1900s, science has promised a machine that would flawlessly translate one language into another. Today computers can transform technical documents in certain languages with up to 80 percent accuracy. Now interest is again building to create machines that will handle all major languages with nearly pluperfect precision.
Japan, as part of its fifth-generation computer project aimed at building ultrahigh-speed machines, is working on an automatic translator with a 100,000 -word vocabulary that will transform Japanese text into other major languages with 90 percent accuracy. The European Community (EC) has a five-year Eurotra plan that seeks to unscramble the seven tongues of its members. Dozens of other research projects are under way around the world.
Spurring the renewed interest are advances in artificial intelligence that promise better-performing and perhaps multilingual machines. There is also a pressing need for new ways to unscramble tongues in today's polyglot, communications-dependent world.
No fewer than 5,000 languages are spoken around the globe. Human translation can be a slow - and expensive - process. Japan spends more than $2 billion a year translating foreign material. As much as one-third of the EC's annual budget goes for similar work.
The quest to create language machines has plenty of antecedents, of course. Some of the earliest work was done at Georgetown University in the late 1950s, where the focus was on turning Soviet scientific literature from Russian into English.
Most of these programs, though workable, were crude at best. Backed by federal funding, the research continued in earnest until 1966, when a National Academy of Sciences panel rejected machine translation as impractical and the government tap was turned off.
Some research continued, however, and the machines, although still rudimentary, developed to the point where they could at least be used to assist professional translators. Today probably half a million pages of text are translated mechanically around the world each year. One of the most ambitious programs is the EC's Systran system, which handles some French-to-English and English-to-Italian translations. Canada uses one to translate French-English weather forecasts. Japan's Kyoto University turns English technical material into Japanese.
In the United States about half a dozen companies now sell machine-translation systems. The latest entry is Challenge Systems Inc. (Richardson, Texas), which came out with an English-Spanish product last month. University of Texas (UT) researchers are expected to hatch a system later this year.
''There has been a resurgence in machine translation in the US and Europe,'' says Jaime Carbonell, an expert at Carnegie-Mellon University.
Machines still lag well behind human translators in the quality of their work , though. Most systems need a human translator to make corrections since, typically, only about 80 percent of the text comes out accurately; 15 percent will need minor corrections while the rest is virtual gibberish. For companies and agencies that deal with large volumes of foreign material, the machines can be timesavers: It takes less time for a translator to clean up sentences then to translate every line.
On the other hand, the devices usually work only for specialized publications , such as scientific, technical, or legal works. These contain words with strict usages and universal meanings. Literary works are far too subtle. Even then, some companies find the technology too primitive: Caterpillar Tractor Company recently tested an English-to-Spanish system but found it as time-consuming to use as human translators.
The reason the chore is so tricky is the complexity of language. First, different words have different meanings (consider the ways ''Johnny blew it'' could be construed). Metaphors muddy things (''he is making strides'' - does that mean progress or walking?). Some words in one language can't be expressed in another.
One way researchers are trying to get around this is to develop software that can analyze semantic structure at the sentence level rather than just the word or phrase level. Others are working on systems that comprehend both sentence structure and syntax.
The perfect translation, of course, will always remain out of reach, even for human beings. But many experts expect, within five to 10 years, to see widespread use of machines that can handle several languages - with little editing needed afterward. ''The state of the art is progressing,'' says Jonathan Slocum, manager of a machine-translation project at the UT's Linguistics Research Center.
Yet others remain dubious. ''In the last 20 years nothing significant has been discovered,'' contends Martin Kay, a scientist at the Xerox Research Center in Palo Alto, Calif.