Argentine company developing leading machine translator

Whatever you do, don't sneeze

Just when Argentina devalued its peso, New Zealander Charles Campbell and a group of local language teachers didn’t hesitate in leaving their academies behind to begin a translation service—and their bet paid off. In 2005, they started Translation Back Office (TBO )with a team of 60 professionals in Cordoba, Argentina to assist clients such as SAP,  Banco Mundial, DiGi-Key Microsoft  and Youtube, among others. In 2011 alone, they processed over 28 million words. Today, they are confident that they will continue to grow in a market worth $8 billion a year. Their calling card: an artificially-intelligent translation machine.

“It’s been a year-and-a-half since we began developing a translator that not only follows rules, but also performs statistical learning,” said Lucas Brizuela, operations manager at TBO. “It learns as it translates and we also add our eight years of experience. This results in a four-fold increase in output from each translator while generating high-quality work.” This counts for a lot as 70% of their budget goes towards paying translators.

“We hope to cut this cost in half, so the translator in a win-win for us,” he continued. “It’s not meant as a translator replacement but rather to make them an editor.” Brizuela hopes to have the machine begin performing English to Spanish and Spanish to English translation this year as that is 85% of TBO’s business. After that, there are plans for English to Portuguese and English to French translations. The next step would be to then focus on non-Latin alphabets within a year or two.

TBO may not be breaking entirely new ground, but it is leading in what will become an industry standard. Machine translators, such as Systran, have existed for years. However, they work via a rules-based system—something that creates great complications when dealing with upgrades of specialized resources.

“For the last five years, as there has been a great increase in processing availability, statistical learning has become a reality in the market with work being reduced by months on top of using smaller groups,” said Laura Alonso Alemany, a linguist with the Natural Language Processing Group at the National University of Cordoba which supported TBO. She points to Google Translate as the best example. “But for a professional to use [Google] Translate is to expose potentially confidential information, and above all, share his or her most valuable capital: their translation experience.”

The brains behind TBO’s machines is Moses, a GPLed translation platform supported by a number of universities and organizations including MIT, Edinburg and DARPA, among others. Other tools that have been incorporated, such as IRSTLM, also come from open source. However, until now these tools have only been used in an academic setting. In order to put the code to work, TBO developed a system of morphological and syntax rules. On top of that, it incorporates style guides, interfaces with industry-standard tools such as Trados and includes a GUI for the translator.