Be a Bugger or be a Thunderbug.
Over the nine years that Language I/O has provided realtime translation solutions for customer support organizations, we’ve learned a few things. We’ve learned that companies are looking for operational efficiencies in customer support. When companies go global, staffing up native speaking support agents for every language is the opposite of an operational efficiency. If companies could empower their monolingual (i.e. English-only-speaking), already-trained customer support agents to chat and email their customers through a translation technology layer, that would be a big efficiency. If companies could spin up a chatbot to deflect those easy-to-answer questions in just one language, and use a translation technology layer to make that chatbot fluent in all languages, they could support those global markets faster and with far less cost.
Why are Language I/O’s Translations the Best on the Market?
What many companies do first is they attempt to build their own translation layer. They create their own integration with a single neural machine translation (NMT) platform like Google, Microsoft or Amazon. At some point they realize that a general NMT platform won’t accurately translate their product names, industry jargon or their customer’s slang. At this point, some will elect to train an NMT model for each language pair they need to support. They’ll realize this is slow going, expensive and requires language expertise they probably don’t have in-house. So they outsource this NMT training work, only to marginally move the needle on quality. Before they know it, a year has passed and it would have been cheaper to stick with native speaking human agents.
We too went down the NMT model training road and ran into the same problem. It’s expensive, time-consuming and not scalable. As the old saying goes, “necessity is the mother of invention” and we invented. Specifically, we invented a way to get our customers – including many Fortune 500 companies – up and running with realtime NMT services that are properly translating messy user-generated content from their customers in a matter of 24 hours. In fact, we’ve been making this business-specific translation functionality available to users of our Salesforce, Zendesk and Oracle Service Cloud customer support translation apps since 2016. Our ability to enforce a company’s preferred translations for problematic terms atop any of the NMT engines we integrate with is a proven technology. It’s tested, hardened and well, one-of-a-kind.
Equip Your Teams with a Unique Self-Improving Solution
Recently, we took this a step further with the development of SIGLO – Our Self-Improving Glossary solution. While our original glossary enforcement technology would load glossaries from our customers along with their preferred translations into our platform and impose those terms across any of the numerous NMT engines we integrate with, SIGLO automatically detects new candidate glossary terms without the customer needing to lift a finger.
In this blog post, I will walk you through a detailed explanation of how Language I/O’s proprietary glossary enforcement and SIGLO technology accurately translates messy, user generated content with lots of company specific terminology, with examples from the gaming industry. While most of our customers are not in gaming, over these past nine years we’ve also learned that gaming content is some of the hardest content to accurately translate. Not only are gamers notorious for flaming one another with interesting, colorful slang and not worrying about grammar – the platforms they use employ a thick dictionary of made-up, fantasy names that have No. Literal. Translation. Let that sink in. Gamer terms like Eldrazi Devastator and Thunderbug were invented by a gaming company. Run these through google translate and you get nonsense.
If Language I/O can handle terms like these, we can handle most tough translation scenarios. Further, we’ve found that other industries have more subtle but equally difficult problems. When it comes to use of a word like “player” in English that could be translated in a variety of different ways into, say, Spanish, a single chat message won’t often convey enough context for even a trained NMT engine to know what type of “player” you’re talking about. If you’re chatting about a basketball player you need the word translated into Spanish as “jugador.” On the other hand, if you’re translating the word “player” in the context of a video streaming platform, you want “tocador.”
Our Solution Explained For Gaming
These are all scenarios Language I/O can handle at a fraction of the time and expense required going the NMT model training route. And here’s how it works.
In the above example we have a Spanish Monster Hunter chat in the middle box. Everything in orange in that box is problematic:
- Mosgorrinos and insectruenos are fantasy game characters – there is no literal translation for them. Monster Hunters has made-up equivalents for each such character and they should be translated into English as mosswines and thunderbugs.
- Cheteados is an adjective that indicates a character’s power level is a bit of an overkill. But to be meaningfully translated into English, it should be “OP” which stands for over powered. Not literal or intuitive.
- Farmear takes the English verb “to farm” and turns it into the unconjugated Spanish verb “farmear.” In reality the Spanish verb “to farm” is cosechar.
On the left you can see how Google translates it. It doesn’t try to do anything with “mosgorrinos,” it incorrectly turns “cheteados” into “checked” and translates “insectruenos” into “buggers.” Despite the fact that Google accurately translated “farmear” into “to farm,”no gamer would understand this translation.
On the right you can see how it is translated when it’s passed through Language I/O. Language I/O imposes the preferred Monster Hunter terms onto whichever engine we choose – which in this case turns out to be the DeepL platform. This is an ideal use case for our APIs. Not just gaming companies, but retail, online meeting and learning platforms and social media can now all access this easy-to-spin-up, accurate translation engine from wherever they write code that can hit our service.
Our Solution Explained For Your Customers
Lastly, let’s look at a customer support example. As I mentioned earlier in this post, Language I/O has ready-to-install apps for realtime chat, email and chatbot translation for Salesforce, Zendesk and Oracle Service Cloud. No coding needed. Here’s a scenario from a Chinese-speaking Magic the Gathering user and an English-speaking customer service agent.
In the above example everything in orange in the middle box displaying the incoming Chinese gamer chat is problematic.
- “毁世奥札奇牌” is another fantasy game card name – there is no literal translation for it. In Magic, the equally-made-up English version is “Eldrazi Devastator.”
- “歼灭” is the description of an action and isn’t entirely literal – should be “annihilator.”
- “瞎菜了” even though literally means something like “blind dishes” is slang for being confused.
- Lastly, “gkd” is a recently invented chinese acronym where each of the latin characters represents a Chinese word and they all equate to “do it now.”
In the image above, on the left you can see how Google translates this Magic support request. It translates the game card name “毁世奥札奇牌” phonetically and incorrectly into “Zhang Zhaozaiqi.” That translation is useless. It comes closer with “歼灭” translating it into “annihilation” which is still wrong however and especially when you’re the support agent and have to search your knowledge base for a trigger term, it needs to be the right term. For “瞎菜了” Google goes with the literal, nonsensical “blind dishes” translation instead of the accurate “I’m confused.” And lastly, Google doesn’t do anything with “gkd” because the chat acronym is too new for Google to understand. Suffice it to say the support agent would provide very poor support with the translation on the left.
As organizations seek operational efficiencies in the realm of product and support globalization, translation technologies are going to be playing a bigger role. The options available today really fall into two camps.
The take-a-year-to-train-your-own-nmt-model-for-every-language camp. Or the Language I/O camp, a.k.a. the be-a-bugger-or-be-a-thunderbug camp. I recommend you be a Thunderbug.