How do you translate an address into 26 languages so that it sounds perfectly natural in each one? London-based startup what3words has given every location on Earth a three-word English address; now, it’s working to support as many languages as possible so that users worldwide can find any destination in their native tongue.
What3words divides the world into a grid of 3m x 3m squares, each of which is assigned a unique three-word identifier. The company’s headquarters near Portobello Market, for example, can be found at “index.home.raft”. Since being founded in 2013, it has raised over £10 million in funding, opened offices in South Africa and Mongolia, and partnered with companies from navigation app Navmii to carmaker Mercedes-Benz. In the United Arab Emirates, Aramex delivery drivers use what3words to take packages to the right house; in Nigeria, the postal service uses it to deliver mail. Order a Domino’s pizza in Sint Maarten in the Caribbean and what3words will help it arrive at your door.
But, as the company continues to expand globally, it faces the challenge of translating a product that is language-specific by definition. “When you're talking about addressing the world, people expect to be able to do that in their own language,” says head of languages Jamie Brown. “It’s really key to our global reach to make sure people have the tools to talk about their own location in the language they're communicating in.”
It’s not as easy as translating addresses word-for-word. The English what3words map is based on 40,000 of the most common words in English, but they may not be the most common words in Swahili. Other problems soon crop up; Brown gives the example of “snowman”, one of the words used in the English version. In French, this would be “bonhomme de neige” – already three words in itself.
So, to translate the product, the company has to start with a fresh list of words each time. It then hires native speakers to go through and whittle the list down to about 25,000 words, which, when combined in trios, is enough to cover all of the world’s landmass (the oceans are mapped only in English).
The language team considers many features in order to select the most appropriate words. “It's really manual,” says Brown. “We believe that languages can't be processed completely with a computer.” Offensive words are out, as are slang terms. The team must also weed out any homophones (words that are pronounced the same), as users have to be able to share addresses verbally, or using voice recognition, without confusion. Crucially, each word must have a unique spelling across all languages to avoid duplication – so “barn”, the Norwegian word for “child”, isn’t used, because it already appears on the English list.
Different languages throw up different linguistic puzzles. French verbs are particularly rich in homophones, with the infinitive form often sounding similar to the “vous” form or the past participle, as in penser/pensez/pensé; what3words uses only the infinitive in these cases but will accept any of the other spellings for it. In Mongolia, people often type Cyrillic using Roman characters, but there is disagreement over whether the Cyrillic “x” should be written as “kh” or “h”, so the map displays the same results regardless which the user opts for. For Thai, the language team had to decide what even counted as a word, as the language is written without any spaces. “Often it's on a word by word basis,” says Brown. “There was a natural difference between ‘computer program,’ which is one concept, and ‘chicken green curry with noodles.’”
The map can still end up with some very similar-sounding addresses, but these are assigned to locations that are purposely a long distance from each other, in order to minimize mix-ups. Strawberry.snowman.rainbow can be found in Namibia, for instance, while strawberries.snowmen.rainbows is in Canada.
The what3words app currently supports 26 languages and is about to launch in China. Chinese presents many complexities for translation, says Brown – not least because it is not one but a collection of languages. His team started from the basis that many Chinese languages use the same written characters, so a word taken out of context can be understood by both a Mandarin speaker and a Cantonese speaker, for example. “They’d pronounce it differently and they'd use it in a sentence differently, but on that basic level the character itself means the same thing,” he explains. When using speech recognition, what3words will accept Pǔtōnghuà pronunciation – a standard variety of Chinese that most people speak in addition to regional variations.
It’s a painstaking process, but it’s important to get right, says co-founder and CEO Chris Sheldrick, as there’s no opportunity to change the word lists at a later date. “When we make a what3words language, it is then fixed forever, because one of the key properties of what3words is that it works offline,” he explains. “You could find a three-word address, write it on a scrap of paper and in ten years’ time you could pick it back up and it’s going to point to exactly the same place.”
Japanese: Japanese is the only what3words language that doesn’t use a full-stop between words, instead opting for an interpunct (which looks like the top dot in a colon).
American English: Words with variant spellings in British and American English are avoided, so you won’t find “color” or “colour” in an address.
French: What3words ignores accents as not everyone types them. This means it can’t use words that are only differentiated by accents, such as côte and côté.
German: A German person using a non-German keyboard may type an umlaut by adding an “e” on to the vowel instead, so that “schön” becomes “schoen”, for example. Both spellings will take you to the same address.
Norwegian/Danish: Norwegian and Danish share lots of words, so these have to be split between the two languages’ word lists; they can’t appear on both.
Arabic: Multiple countries have Arabic as their first language, and some words that are common in one place are slang in another. Consultants are taught to filter these out.
Chinese: A Chinese what3words address can have multiple characters in each of its three parts – meaning each one is technically considered a “phrase”, not a word.
This article was originally published by WIRED UK