Machine Intelligence: Case Study
One of the world’s largest software providers came to us for help with their internet-based maps. Used by billions of users globally, these maps display the names of the billions of locations around the world. From continents, countries, and names of capitals down to names of the smallest villages and creeks in every country.
This software company wanted to take all the map locations around the world and localize them from and into several languages so each user could read location names in their own language and writing system. Easy to do with a team of human translators...but very expensive and time consuming.
Conventional machine translation solutions can provide a cheaper alternative to human localization, but they are not designed to deal with map localization conventions in an appropriate way. This can result in low quality localization.
Localizing map locations is similar to other proper name localization where, for example, “Miller” in “John Miller” is not translated into the Russian or Chinese words for “miller.” Instead it’s written to allow a Russian or Chinese speaker to pronounce the name similar to how “Miller” is pronounced. This type of conversion where the pronunciation rather than the word meaning is preserved is called transliteration.
We developed a machine learning solution, a transliteration engine that took map localization conventions into account and enabled high-quality automatic localization of map locations. The transliteration engine enabled us to limit the role of the human translators to just review and correct the machine output, instead of doing the localization manually.
In automatic evaluation we achieved between 41-77% completely correct transliterations, when compared to human “gold” standard. But manual review of these figures revealed that the majority of the “errors” could actually be considered acceptable variants.
Overall, the introduction of the transliteration engine into the map localization workflow increased the throughput of localized terms significantly. This lead to cost savings per localized entity by turning what before was a pure localization task to one that was simply a review process. The result was faster scaling to new languages and locations.