Tilde, based in Riga, Latvia, is a leading developer of European translation technology. The company focuses on developing hybrid machine/human translation engines for secondary languages, and started with the Baltic languages (Latvian, Estonian, and Lithuanian), which are poorly served by mainstream translation platforms and services.

Let’s MT, one of the company’s anchor products, is a custom machine translation engine that enables customers to build their own machine translation platform.

Statistical machine translation engines are trained using sets of aligned texts, where each sentence from the source language is paired with a high quality human translation. Statistical MT engines all use the same underlying algorithms (most are derived to some extent from an open source project called Moses). Translation quality is primarily determined by the quality and size of the training sets fed to the engine.

Tilde has developed a set of admin tools that enable less technical users to create, train and deploy their own custom translation engines (building your own Moses server is otherwise very difficult to do). With their tool, you simply upload your training documents, and can also used shared training data compiled by Tilde, and then deploy your translation engines in the cloud. You also can test drive their Baltic translation engine at translate.tilde.com

The company has amassed an impressive collection of training texts, with 1.4 billion translated sentences spanning 102 languages, that has been used to train 129 separate machine translation systems. This has led to significant productivity improvements, by lessening the requirement for post-editing, with savings ranging from 25 to 30% for languages such as Czech, Polish and Latvian.

Best Fit

Companies that need to incorporate machine translation, either as a stand-alone service, or for first-pass automatic translation followed by human post-editors, and who need this in languages that are not well covered by mainstream translation engines. Companies that also have a large set of training data that want to create a custom translation engine, for a specific language pair or subject area, will also benefit from tools like this.