The Art of Translation
Ball, Stefan, Contemporary Review
For years beleaguered humanists have been arguing that artificial intelligence is impossible by defining intelligence as anything that a computer can't yet do. The parallel argument in the translation world would be that literature is anything that a machine cannot translate.
`Anyone who tries to use machine translation to translate a literary work is wasting his time', says John Gregory of BG Systems, a company which specialises in providing computer solutions to translators. If this is so, how can one explain the claim at a recent conference that some three hundred million words are translated by machine every year?
The clue lies in the kind of texts that are successfully translated-and the kind of translation the machines provide. `The main successes have come in small subject areas, like the Canadian METEO system', says Gregory. In 1990 METEO, a custom-built computer program, was translating some 45,000 words' worth of weather bulletins from French into English every day. Weather reports, with fairly well-defined technical terms used in very specific contexts, are ideal for a relatively simple, dictionary-based system. To produce an adequate translation the computer does not have to know about anything except the way in which words dealing with the weather are put together in two different languages.
The US Air Force has also made use of machine translation to produce quick, rough drafts of technical and scientific papers. But once outside the restricted domain of a single subject area, the results are less satisfactory, and if a paper is found to be worth further study it is passed to a human translator who produces the finished text. After all, would you like to read more than a page or two of raw machine translation like this one, produced by the European Union's SYSTRAN system: `But there him remained to achieve a difficult acclimatization period, for of carbon producer it had become steel user?'
This is a fairly good translation by machine standards, where reduced quality is the price paid for speed: the United Nations target for a professional translator is 2,000 words a day, but the larger computer systems think nothing of ploughing through 700 pages in a night. At the same time the translation world is full of tales of mistranslation, with `company posts sizable growth' coming out as `guests mail large tumour' or `late July' being translated into `dead July'. Apocryphal or not, stories like this highlight the fact that machine translation systems make stupid mistakes because they don't understand a single word that they are translating.
In the end humans are almost invariably involved in the machine translation process, either through pre-editing (rewriting the `input text' to make it simpler and grammatically orthodox) or post-editing (rewriting the machine output to make it more readable). Post-editing of the SYSTRAN translation, for example, gave: `But he had a difficult adjustment period ahead, for from being a coal producer he had become a steel consumer'. Both pre- and post-editing have to be done by a translator, since reference has to be made to the source text: just because the translated text makes sense, it does not mean that it necessarily makes the sense that was intended by the original author.
It can be seen then that the claims made for machine translation by the people who are trying to sell it are at best exaggerated. Talk of 90 per cent accuracy is meaningless when no one is sure what is being counted. If it simply means that one word in ten will be wrong or out of place the experienced human translator may find it quicker to ignore the machine-produced version and start over again with the source text, rather than trying to decide which of the translated words are wrong.
Literary texts are of course as far away from restricted domain texts as it is possible to get. Not only can any kind of vocabulary be employed to talk about any kind of subject, but there is often deliberate skewing of words so as to produce new collocations and metaphors, along with the constant use of ambiguity and irony so as to make the text work on more than one level. …