Machine translation for a marketplace
The business goal
The business goal of this project was to build an automated translation service for the retail website that specializes on trading Japanese car parts. We needed to create ad text translators from Japanese into Russian. For this purpose, we chose a statistical translation model based on a comparison of a large body of parallel texts. They were used as client texts received from the Q&A service (questions in Russian and translated into Japanese).
From the technical point of view, we used several statistical translators such as Travatar and Moses. We then used Travatar, since it showed the greater quality translation based on objective metrics. One of the key challenges that we faced is the low quality of the translation and lack of alignment at the level of sentences. To align it we developed a statistical algorithm, that worked by searching the closest statements on the basis of the dictionary of n-grams. We used Python and KyTea and EDa for this purpose.
The project resulted in a sucсessful implementation of translation services into our client’s website. Within a couple of months, our client saw a rise in sales of over 20%.