Linguistic skeptics like Alan Melby have explained why a computer can’t translate like a person. Developers of machine translation (MT) have largely conceded that point and focused on developing systems that provide rough information (“gisting”). But if MT is still unable to make a contextual distinction between two teenagers and an optical instrument, can it ever really be useful?

It turns out that it can, but only if it ceases to be purely machine translation. It’s as if developers said, “You know, maybe Melby has a point. Maybe language is just too complex for us to ever develop good enough rules to make machine translation really useful for general purposes. But can we use computers to capitalize on the linguistic creativity of human beings?” This was the starting point of academic research into what was initially known as example–based machine translation (EBMT). Rather than struggle to make ever more complex rules systems for analyzing the source language and transforming it back into the target language with the aid of a dictionary, why not use a computer to sift through a corpus of human translations and pick the best matches for a given sentence to be translated?

The EBMT approach did not make the jump from research institutions to practical application in its first incarnation. It was just too difficult for a machine in those days to come up with enough examples and analyze them sufficiently – unless the “machine” was a human being. If a computer could show a person examples of how a sentence (or very similar sentence) had been previously translated, the person could use his own linguistic skills to choose the best example, or modify it to fit the new sentence. This approach, now known as translation memory (TM), has revolutionized the field of translation in the past ten years.

TM can speed up the translation process and enhance consistency with minimal loss of quality, if used correctly. However, its speed is still limited by the length of time a person can work with full attentiveness, and its quality by the skill of the translator. And because a TM system has no linguistic intelligence of its own, it only works at all if there is a human translator available for the desired language pair – a real problem for many languages. Finally, a TM system breaks the source text into segments – usually entire sentences – and checks them against the existing translated segments in the memory. It is generally not able to compare smaller phrases inside one segment to phrases of other segments and suggest translations. Researchers, such as those at the French firm Lingua et Machine (developers of the Similis TM system), are working on “second–generation” TM systems. The T! AUS (Translation Automation Users Society) has begun referring to early commercial tools in this area as “Advanced Leveraging”. Even as these early tools are hitting the market, though, they may be superseded by more powerful technology.

The enormous speed of modern massively parallel computing, combined with the staggering amount of translated content now available on millions of websites, has revived the seemingly lost cause of EBMT, in a much more sophisticated form referred to as Statistically Based Machine Translation (SMT). The huge advantage of the SMT method is that the machine no longer has to “know” the context to decide, for instance, what “jumelles” means. It analyzes the collective wisdom of a huge database of human translations, assesses the probabilities of the alternative translations and incorporates the most likely candidate into the translation. It’s a fairly safe bet that, with a sufficiently large corpus of examples to analyze, the statistical process would generate the correct translation of “jumelles,” because there are a lot more sentences in the real world that refer to 18–year old twi! ns than to 18–year–old binoculars.

If SMT technology lives up to its auspicious beginnings, it may have sweeping effects on the language industry, not least on McElroy Translation. Executive strategy at McElroy embraces the potential of this dramatic MT technology advancement. As was the case with translation memory technology, the inclusion of SMT will open up new types of translation work that were never before feasible. For some large localization projects for instance, a judicious mixture of TM and MT (“MTM”) can lead to reduced cycle times and greater productivity. These are exciting times for the translation industry. The best part is that instead of threatening the value of human translators, these new technologies increase it.