Need to Translate Sloppy Content? No problem.

A few months ago, I pitched Language I/O’s new approach to instantaneous translation of customer support content to a sea of translation tech professionals and a panel of judges. We were at the LocWorld Process Innovation Challenge in Silicon Valley. The other presenters and I had been selected from a larger pool of contestants. For months we had been competing for this coveted opportunity to showcase our tech in front of the very companies poised to use it. As finalists, we had to box everything we needed to say into four, tiny minutes.

YouTube player

Now it’s time to zoom into the inner-workings of the Language I/O machine learning (ML) approach to realtime translation and how it constitutes a breakthrough for the customer support industry. Actually, it constitutes a breakthrough for any industry that needs quality translation of messy, user-generated chit chat.  

When a Fortune 500 company is taking its customer support organization global, there are really two options for providing foreign-language speaking customers with chat and ticket support.

Option 1

They can hire native-speaking support agents. If you look at the cost of a support agent globally – averaging the more expensive salaries in Tokyo in with the crazy low rates in the Philippines – it costs about $40,000 USD annually per agent. Even for business-hours-only support, they’ll need more than one agent per additional language and most large companies support at least ten languages. For your new, international support team, annual costs quickly rocket to over $1M. 

Option 2

Owed to the hefty price-tag attached to hiring native-speakers, large companies are recognizing that it’s far less expensive to let existing, monolingual support agents use Option 2: Language I/O. 

Our tech plugs right into Salesforce, Oracle and Zendesk – or the customer relationship management (CRM) platform where support agents already work. In a nutshell, Language I/O empowers existing, English-speaking (or monolingual) employees to chat and email with customers in any language. We do this via a hybrid of rapid-turn-around human translation services and neural machine translation (NMT) engines. But I’ll be honest. Given the higher cost of human translation vs. the low, low cost of the ever-improving NMT, our customers are moving fast towards pure NMT for translation of chats, emails and even articles or FAQs.

Language I/O + Machine Translation

Early on in our business, we considered building our own NMT solution. We could have started with the amazing OpenNMT provided by Systran and continued building from there, but Systran already does a great job. As does Google. As does Microsoft. As does DeepL. So why reinvent the wheel? Our strategy? Use ALL OF THE NMT ENGINES. That’s right, we’ve created integrations with the best NMT engines in the world in order to provide the best NMT service in the world for customer support translation. Which brings me to the Language I/O value-add in this space:

A surprisingly large number of Language I/O customers come to us after they have tried to code their own integrations with one of these NMT engines for chat or email translation. Every time, they find that they have many, many company-specific terms that a general NMT engine won’t properly translate. They could train these engines to properly translate their specific terminology, but when you support 10 or more languages and have to train per language pair, that’s 20 trainings! For each of these 20 trainings, you’d need 20,000 or more human translated segments just to get started and you’d have to continue to maintain and test the trained engines over time, or the quality degrades. Even enterprise customers with deep pockets are just saying no.

At Language I/O, we recognized that per-language-pair training is not cost effective. In its place, we have developed a machine learning strategy that sits on top of our numerous NMT integrations for a unique solution to the training problem. 

First, we recognize that certain terms simply require company-specific translations. General NMT engines don’t have the context required to accurately translate them. Take the example of a support chat that came through from a Spanish-speaking user of an online betting platform. 

Let’s ignore the missing accents and lack of punctuation. In this 19-word Spanish chat, there are three misspellings and various other oddities–all highlighted in orange. Some NMT engines will fix misspellings without help. Others won’t. 

  • “ganafor” is a misspelling of “ganador” or in English, “winner”
  • “dafo” is a misspelling of “dado” or in English “given”
  • “iltima” is a misspelling of “ultima” or in English, “latest”

Our Spanish gambler randomly used one English word (“free”) and two tennis-specific abbreviations (rafa and rg). Untrained NMT engines always lack the context required to get these abbreviations right. 

  • “rafa” is a nickname for Spanish tennis pro Rafael Nadal
  • “rg” is an abbreviation for “Roland Garros Tennis Tournament,” which in the US we know as the French Open

If the company that received this translation were to run this chat through general NMT via a home-grown integration, the chat would have been translated to something like the below. 

In this case, the English-speaking support agent would be stumped. 

Language I/O leverages both rules-based and machine-learning decisioning processes (which some call “AI” for artificial intelligence) to give the unsuspecting support agent a translation she/he will understand.  First, when we onboard a new company, we work with them to identify terms like “rafa” and “rg,” which require special handling. These terms and the company’s preferred translations for them, are quickly loaded into our native glossary. Then, Language I/O simply forces the NMT engines to use these company-specific translations whenever the terms are recognized. There’s no formal training and no loading of 20,000 human-translated segments per language pair. 

Next, comes our actual machine learning. Run the messy chat from our Spanish gambler through five different online NMT engines and you’ll get five different translations. Some engines will fix the grammar and spelling issues for you. Some engines translate certain language pairs more fluently than other engines. It varies day-to-day because these NMT engines are constantly learning and adapting. 

The Language I/O machine leaning model predicts which of our integrated NMT engines is going to best translate the chat. Or the email. Or the social media post. Language I/O makes this prediction before it sends a translation request to the best NMT engine for the message translation. How? Our platform performs actual machine learning. It learns from feedback loops containing helpful data that we’ve been piping into our database for years. 

Some of the more interesting feedback sources include edit distance data from our human translation platform. We also learn from customer satisfaction (CSAT) scores from our integrated CRM platforms where these chats and emails come from. Remember, our solution plugs directly into Salesforce, Oracle Service Cloud, Zendesk and other CRMs. We have access to customer satisfaction scores for every messy chat and error-ridden email we translate. That data is used to tell our machine learning engine how well it’s doing and how to improve its quality prediction capabilities. 

Language I/O randomly select chats and emails that our engine has already sent to an NMT engine for translation. Our platform then passes these sampled messages back in for translation but this time through all of our integrated NMT engines. Language I/O then pipes the comparison machine translations directly to our human linguists for a quick post edit. 

The goal is to see how much the linguist has to change each of the machine translations to make each human-translation-quality. This process produces a number called an “edit distance” score. The lower the edit distance, the less a linguist had to tweak the machine translation to make it human quality and hence, the higher the original NMT quality. We compare these human-generated edit distance scores with the score generated by the Language I/O machine learning platform. In this fashion, the Language I/O machine learning model learns and all by itself improves its ability to predict which of our integrated NMT engines will best translate a support message the moment it hits our system for translation.

Take a look at the image below and compare the raw NMT output to Language I/O’s translation.

Attempting to explain this incredible technology during my time window at LocWorld felt silly. The audience did laugh at the horrible raw NMT translation, but nowadays everyone says their tech is “AI powered” so I saw eyes gloss over with skepticism when, during the period of a single minute, I dove into our tech. I’m positive that many were thinking “yeah, yeah–I bet it includes blockchain too.” In the end I placed second out of the bunch. I was pleased, given the difficulty of explaining a machine learning model  in four, tiny minutes. But watch out world – this isn’t tiny. It’s big.