Salesforce A.I. researchers came up with a faster way to translate text -- and it's based on tech from Google

Key Points
  • The Salesforce researchers are building on technology that Googlers unveiled earlier this year.
  • The technology isn't being used in Sales Cloud or other available Salesforce apps, but it could eventually help the company deliver documentation in different languages.
From left, Salesforce intern Jiatao Gu, chief scientist Richard Socher and research scientist James Bradbury.
Jordan Novet | CNBC

Salesforce on Tuesday revealed its latest progress around ways to translate text from one language into another without human input.

The translation system comes close to being the most accurate there is today by one measurement, and the researchers have ideas about how to improve it.

"We want to see it deployed in a lot of places," Salesforce chief scientist Richard Socher told CNBC in an interview at the company's Salesforce IQ office in Palo Alto, California.

In recent years, web companies like Facebook and Google have been the ones leading the way in terms of incorporating artificial intelligence into their products. Now Salesforce is among the companies drawing on AI to improve the software that people at big companies use. That could help Salesforce as it competes with the likes of IBM, Microsoft and Oracle.

Salesforce established its AI research lab following its 2016 acquisition of Socher's start-up, MetaMind. The group has doubled in size since then, Socher said. Like many other corporate research organizations, Salesforce Research regularly publishes results of its research and releases code online for free for others to explore.

In this case, Salesforce circumvented technology called long short-term memory networks (LSTMs) that have become popular for machine translation. Instead, they chose to build on a system called Transformer, which Google unveiled earlier this year.

Translation has been revolutionized by the use of a trendy type of AI called deep learning, which involves training systems called artificial neural networks on lots of data and then getting them to make inferences about new data.

But the neural network-based systems can be slower to make inferences than more traditional architectures, as Socher and four of his collaborators -- Jiatao Gu, James Bradbury, Caiming Xiong and Victor O.K. Li -- point out in a new academic paper. Simply put, their model comes up with the best words for a translation in parallel, instead of in sequential order, which provides a speed boost.

Bradbury explained some of the team's work with an analogy.

"You have a whole panel of professional translators, and they're all in separate cubicles somewhere," he said. "You want to get a translated sentence, but you don't let them talk to each other. When you ask each of them to output one word, they're not going to make the right guesses about what each other is thinking. They're going to produce something that totally is gibberish. That's the problem that we ended up solving."

While Salesforce already offers technology that's based on work from its research lab, machine translation isn't in use in Sales Cloud, Service Cloud or any other product. That could change in the next few years, though.

"I think we'll see especially for technical documentation" that it can come in handy, Socher said.