Network: A way through the European Babel

The European Commission has long felt the need for translation software. Back in 1976, it bought rights to Systran from the WTC (World Translation Centre), a company owned by the idealistic Hungarian-born American inventor Peter Toma, a veteran of the Second World War, who saw machine translation as a contribution to world peace. Toma worked on a number of prototype systems before the US Air Force bought a Russian-English module in 1974 and the Commission a prototype French-English language pair in 1976.

Inside or outside the Commission, it never instantly impressed. The first standard vocabularies were useless for bureaucratic purposes and the output was painfully slow. In 1976, a BBC film crew rushed away from the EC's development centre to catch the last ferry from Calais after waiting more than five hours to film one punch-carded paragraph of translation emerging miraculously from the machine. Even as late as 1988, only 4,000 pages were produced. But by 1995, the figure had risen to 170,000. Response time is now down to six minutes. Systran has become an essential part of the service accessible by E-mail to every PC in the Commission.

Initially a bilingual system, Systran now adopts a modular approach by which one language analysis module serves for multiple target languages. Seventeen language pairs exist, from four main source languages. Language pairs are improved gradually. English-French, French-English are the most acceptable after 20 years of development. The system uses a "bottom-up" dictionary system rather than attempting "top-down" grammatical analysis. If a word is ambiguous, Systran will always throw out the most frequently occurring option.

In the beginning, the technique was inevitably fallible, as in the French- English translation of the phrase "vis-a-vis", which defaulted to the English axiom "live to screw". As new domain codes and extensions to the phrasal dictionary rapidly improved performance, its sense of humour diminished.

So how good a translator is Systran? While it still outperforms newer, statistically based translation systems, it cannot cope close to human performance. In practice, Systran is used for quick translations of temporary or internal documents, each page carrying the periodic warning "!!RAW MACHINE TRANSLATION!!". Any legal or externally published documents are still put in the hands of human translators.

The new strategy is to promote Systran throughout all institutions of the Union rather than simply among translators. There are an estimated 2,000 users who can exchange a document in a foreign language for an approximate rendering in their mother tongue. This demotion has led to its greater acceptance, as it is no longer seen as a threat to jobs.