Sign In

Communications of the ACM

ACM TechNews

Salesforce AI Researchers Came ­p With a Faster Way to Translate Text--and It's Based on Tech From Google

Input architecture for the Weighted Transformer for machine translation.

Salesforce artificial intelligence researchers are using Google's Transformer system to translate text without human input.

Credit: Karim Ahmed, Nitish Shirish Keskar, and Richard Socher

Salesforce artificial intelligence researchers on Tuesday disclosed details of their project to translate text without human input, using Google's Transformer system unveiled earlier this year.

Instead of determining the best words for a translation in sequential order, the Salesforce team says its model performs this task in parallel, thus accelerating the speed of translation.

The researchers say their system comes close to being the most accurate currently in existence by one metric.

Salesforce researcher James Bradbury notes the traditional technology of long short-term memory networks is akin to keeping translators separate and isolated while they translate a sentence, without letting them communicate with 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," Bradbury says. "They're going to produce something that totally is gibberish."

He says Salesforce's new method addresses this problem.

View Full Article


Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA


No entries found