Machine Translation – What You Need To Know
The new Czech government faces many difficult decisions. For example: ‘Will from next year to ride with a plastic box behind the windscreen, or a continuation of classical vignettes?’ So, at least, informs us Google’s machine translation of Lidové Noviny.
This is what you get for relying on free machine translation. More sophisticated and expensive translation software is available. If you spend a great deal of time preparing your documents in a form the computer can understand, such translation produces something quite passable.
Translation requires the translator to know what the passage means. Philosophers might dispute in what sense a computer ‘knows’ anything: certainly, however, it knows only a part of the meaning of a text. A message’s meaning is determined to a large extent by its context – a context which a computer’s capacity for machine translation simply doesn’t have.
Susumu Kuno, a researcher at Harvard in the 1960s, once asked a computer to process the sentence ‘Time flies like an arrow’. Everyone knows what this means. But the computer wasn’t sure – it found various interpretations. Perhaps it meant that a particular species of flies – time-flies – take pleasure in an arrow? Or was it a command to time a fly like you’d time an arrow? We understand the meaning because we know that ‘Doesn’t time fly?’ is a common sentiment, whereas timing arrows doesn’t happen outside the archery butts and timing flies is almost inconceivable. But a computer doesn’t have this social background.
In the opening paragraph of this article, Google’s machine translation was confused by the word jezdit, which can mean ride, but which also means to go by any form of transportation (rather than on foot). Since we know that most people don’t travel on horseback, we also know that ride is very unlikely to be the correct translation. But Google doesn’t know that. The interpretation of vin?ta – Google’s ‘vignette’ – shows why you need not just human translators, but human translators versed in the relevant culture. It’s a token showing that drivers have paid motorway tolls.
In many cases we simply don’t understand how the brain is able to make sense of a passage of text, and if we don’t understand a process in our own minds, we can hardly get a computer to simulate it. How – fundamentally – do words get their meanings? Can the words part and whole be defined other than by reference to each other – and if not, doesn’t that give us a circular definition? If a girl is a young female and a boy a young male, why can females still be girls into their twenties when boys generally can’t?
All of this presents a fascinating field of philosophical inquiry. It’s no good, however, for the programmers of machine translation software, who need a coherent theory of meaning that they can transform into algorithms for the software. It’s actually possible that one day automated translation software will work – but that day will be a long time coming.