Language algorithm that translates into sales
Start-up of the week
Published 13/03/2014 | 02:30
So you have a start-up. You have the backing, the people and a few customers. Now what? For many, export is the name of the game. But to do that, you may need your service to be available in a different language. Welcome to world of 'localisation', a process that everyone from social networks to washing machine manufacturers needs.
Localisation is expensive. And this is where Dubliner Tony O'Dowd, right, comes in. His startup, KantanMT, is an online 'translation engine' that offers high accuracy translations for technical manuals and other specialist company literature. But doesn't Google Translate do this?
"Yes, but it's very inaccurate," said Mr O'Dowd. "For example, there's an independent translation industry standard called Bleu. Google Translate would score about 15pc accuracy. Our engine scores between 70pc and 80pc."
What KantanMT does is to use advanced prediction mathematics to analyse, say, a technical manual and use it to predict an accurate translation for another in the same genre. For companies producing the same kinds of products, it's a very useful tool. Typically it's best used for technical, medical or academic content," said Mr O'Dowd. "This is because this kind of content is written in a structured, repetitive way.
"Because our system is a self-learning algorithm, it gets to know the structures well."
But it doesn't work for any kind of corporate text.
"Marketing text is very difficult because it tends to be emotive and unstructured," he said.
The system isn't foolproof: an 80pc accuracy rate still requires someone to bring it to 100pc. But the point is to reduce the cost of translation: a 20pc shortfall is a lot cheaper to perfect than 100pc from scratch.
"If you look at one of our clients, MAN Trucks, they have professional localisation experts that typically translate 350 words per hour. Using our system, they can translate 550 words per hour."
The localisation market is a big one. Most estimates measure it at around €20bn in 2013. KantanMT, which is funded by venture capitalist firm Delta Partners and Enterprise Ireland, has been up and running for just over a year, said Mr O'Dowd.
"We have 10 people employed. In the 13 months we've been in business, we've translated 347 million words and seen 26 billion words uploaded by our clients to train our translation engines."
Mr O'Dowd is no stranger to the localisation industry. His previous firm, Alchemy Software was sold to a larger company, Translations.com, in 2008.