One of the most significant developments in linguistic processing has been the expansion and improvement of translation software. Whereas it was once the case that writers had to hire expensive professionals to manually translate even the most mundane of documents, now there are a host of computer-based translation solutions.
Back in the 1980s, when one of the more advanced work tools we had at our disposal was an electronic typewriter, translation software was rare and highly specialized. At large corporations, they may have used mainframe systems such as Systran or Logos, but for the humble home user, the development of translation software was stymied by the processing power required to match speech and translate radically different alphabets, grammar, and sentence structure.
The early pioneers were the fifth-generation computer programmers in Japan who developed parallel processing and logic programming, significantly increasing the capacity of desktop systems. These developments were inspired by the need to translate between Japanese and English for business purposes. Familiar corporate names that required large translation volumes included Fujitsu, Mitsubishi, Toshiba and Hitachi.
Intergraph and Systran began to offer PC translation systems at the turn of the decade, and the AltaVista Babel Fish online translator (powered by Systran technology) went online in 1997.
Meanwhile, in Germany, Trados’s Jochen Hummel and Iko Knyphausen had been working together on translation solutions since 1984. In 1992, they launched Multiterm, a terminology database for translators. As with most systems at this point, what was being offered was an intelligent database, rather than any higher-level logical analysis.
In 1997, Trados GmbH received a major boost when Microsoft decided to adopt its technology for its internal localization requirements. Since then, the company has grown from strength to strength and today offers its SDL Trados Studio package to 270,000 translation professionals.
Translations Memory Platforms
The development of translation memory systems was another huge leap forward. Speeding up the process of translating formal documents such as legal, medical and financial documents, a TMS creates a database of frequently-used terms and calls these up when a similar phrase is used, ensuring that tried and tested, accurate translations are produced.
Such systems “learn” only in the sense that they grow alongside user competence, saving phrases and terminology which have been successful in the past. They cannot easily translate stylistic nuances or idioms.
SDL Trados moved quickly to incorporate translation memories into its computer-assisted translation (CAT) systems. Other key providers of such software include WordFast (established in 2006) and, in the online realm, Google Translate became ubiquitous (other platforms like Gengo and Ackuna have popped up too).
The benefit of such systems moving to Internet was, of course, to draw upon ever-larger corpuses of “translation units,” pairs of phrases recognized as functional equivalents.
With the advent of Machine Translation (MT), improved logic and specified context were added to the arsenal of techniques used to obtain a reliable translation. Word substitutions would be tailored to specific milieus and ever more intricate sets of linguistic rules could be applied to ensure that accurate idiomatic translations were produced.
Grammar-based translation, which is comparatively inflexible and rule based, was gradually replaced by a more statistical model. However, a hybrid system, combining both methodologies, could produce ever more accurate translations.
Artificial Intelligence Engines
In the twenty-first century, voice recognition systems, such as Apple’s Siri virtual assistant, use speech recognition algorithms to identify spoken language and turn it into text. They do this by learning the idiolect (unique manner of speaking) of the user.
Neural Machine Translation (NMT) systems similarly “learn” from individual pieces of text and integrate it all through artificial neural networks to build a statistical model, from which they offer ever more accurate translations. They create a virtual neural network which is intelligent enough to focus attention on different parts of a sentence to create an overall picture of its meaning, rather than translate one unit at a time.
These systems aim towards understanding both the grammar of a language and the stylistics of its individual users. Google is currently leading the way, though there are other companies such as Microsoft, Amazon and DeepL competing to offer the most humanlike translations, which can then be polished up through the post-editing work of a human linguist.
With voice recognition as prevalent as it has become, and translation being increasingly essential across languages with significant orthographic differences (e.g., English, Mandarin and Russian), AI-based neural networks will continue to be a significant area of research in the decades to come.