Cyborgs in the CMS — Using Google’s Neural Networks to Help Real Editors Understand Every Piece of VICE Content

The Challenge

Every day, over 500 VICE editors around the world publish hundreds of articles in 18 languages. Our publishing platform— Vx — is purpose-built to handle that heft quietly and deftly, day in and day out. But earlier this year, as the catalogue and through-put grew, we identified an opportunity to open up a huge new content library to every VICE editor around the world…the one we were already creating.

VICE has such a wide-reaching international editorial team that awareness of new content between different languages and regions is sometimes challenging. Backstage launched with a “worldwide view,” meant to organize everything published everywhere in real time.

But huge growth meant pressure rose on that content fire-hose, and our editors found it harder and harder to keep up — and worse — in research sessions they reported experiencing a sort of editorial FOMO, as more and more articles that were performing well sailed by in a growing number of languages — leaving them unable to consume and consider for localization in their market.

We realized that even the most multilingual editor could only really get their mind across a small percentage of what VICE produced and published each day. Imagine the feeling of scrolling through Reddit, able to see 100% of the thumbnails but only able to read 1-in-5 links — you get a sense of what our “worldwide view” had become for most editors.

We sketched out a few parameters for a new CMS feature that editors could wield to unwind our growing Tower of Babel.

Key Parameters for Interface Translations

1. Expand each editor’s decision set of articles to translate or adapt for their readers to include everything VICE produces in every market.
2. Focus on putting a tool in human hands — allow our editors to bring cultural nuance, local slang and vernacular, and even heavy adaptation to translated articles — no copy-paste machine translations.
3. Make the experience feel auto-magical for the editorial teams.
4. Use the very best technology available.

Parameter 4 meant we immediately reached out to our partners at Google to talk about integration with Google Translate. Since late 2016, with the release of Neural Network-powered translations for key languages, Google has gotten uncanny-good at turning one language into another.

Google’s Apoorv Saxena, who works on the Cloud Translation Solutions team, summarized our opportunity— and our work:

“It’s clear that Vice had a tough problem to solve given the diversity of content and the number of languages they had to deal with. So it’s really great to see the smart, novel approach they took in using the Google Cloud Translation API to extend the capabilities of their editors to benefit their audience with a richer array of stories that might have been missed because of a language barrier.”

V1 — Launch of our ‘GoogleFish’ Feature

One toggle and many languages become readable.

What does the new feature do?

1. Reads the editor’s native language from browser settings
2. Sends strings and requests directly to the Google Translate API, using a technique for batching translations for the fastest performance.
3. Those translations are focused on the list views of our articles first and foremost — effectively turning the headlines of every article in the worldwide view into the users’ native language.

Once an editor has chosen an article they’d like to fork and adapt for their market, we take them here:

The CMS loads the full article preview in native language, and then allows the editor to preview in their native language. This helps them quickly decide if they’d like to fork it for human translation/adaptation before publishing in their market.

What happens next:

1. From this fork-view, an editor can quickly preview the article in their own language to determine if they would like to select it to be translated and adapted by someone on their team.
2. The process turned out to be so seamless, we had to add a yellow highlight to text that had been auto-translated, so that editors knew when source text started in another language, and needed to be proceeded by a human local editor.

In essence, we’ve made every VICE editor octo-lingual — and unlocked the full breadth of VICE storytelling to every VICE editor, everywhere.

What’s next?

Now that we are live, we’re reviewing with our global teams and hungrily gathering data about usage. We store all information related to how an article is created, forked, translated, forked again etc., so we are beginning to see trends not just in global content flow, but in global cultural trends as well.

Along with our own data analysis, we’re also working with Google to extend the concept of “translated pairs,” and begin to push our post-human-translated text back to the Neural Network to improve it over time.