Google Translate is one of the most popular instant translation systems available online, and while it is certainly a type of “machine translation,” it’s quite a different tool than those used in certain situations by professional language service providers (LSPs)such as Skrivanek.
To generate translations, Google Translate (GT) searches millions of sentences for comparable patterns in origin- and target-language documents that have already been translated by human translators and entered into its database. Then, basically, it makes an “educated” guess as to what an appropriate translation would be. This process of seeking patterns in large amounts of text is called “statistical machine translation” (SMT).
You’ve probably seen how GT works: type in words and you will receive a quick translation (in any of 80 languages) that will range in quality from excellent to questionable, depending on how much text for your language pairing has been fed into the GT database. Google Translate director, German computer scientist Franz Josef Och, describes the GT process as the computation of “probabilities of translation” through comparison of the submitted text with billions of words of “learned” text in GT. The more text is available in the database, the “smarter” GT becomes. Tellingly, the GT creative team is made up of mathematicians and programmers and does not include any linguists.*
On the other hand, software systems such as PROMT, Asia Online, SYSTRAN and Moses, referred to as Machine Translation (MT), are complex, customizable translation engines that are specifically trained for certain projects or content in order to maximize efficiency and accuracy. Often used for technical and repetitive texts without subtleties, MT can assist large corporations in the translation of materials they simply would not have the capacity or budget for otherwise.
In the past MT systems were often entirely “rules based” (RBMT), meaning that information about language structures – not mathematical formulas – formed the foundation of their programming. Now MT engines like those mentioned above are often hybrid systems that combine RBMT and SMT. Basically, MT engineers “train” the sophisticated MT programs with glossaries from relevant fields, along with text from specific documents and corrections from previous mistakes, resulting in a tool that becomes more refined the more it is used for each client. This kind of multi-faceted MT requires extremely high levels of capital investment for both hardware and software, and for the process of customization.
Instant online translation tools like Google Translate are a gift in an era of communication expansion so extensive that a large American corporation might want immediate access to comments tweeted by an Icelandic teenager about its latest product. There are numerous instances of such social and commercial interaction online when communication speed is more important than language precision.
But for linguistically and culturally accurate translations of text that contains any ambiguities, nuances or critical information, hands-on human intelligence is still essential. Even complex MT systems are most appropriate for only some types of texts and then merely as producers of raw output that is checked, smoothed and corrected by human post-editors.
*”Google Translate Has Ambitious Goals for Machine Translation,” by Thomas Schulz, Spiegel Online, Spiegel.de, September 2013