Using Machine Translation to Speed Up Localization
What I did
Project
Role: Content strategist, lead copywriter Google I/O, localization lead
Company: Google
Year: 2023
Problem
During Google I/O we often have to accommodate last-minute changes across event content, website edits, and filming captions. When that happens it shortens the time available for localization.
Solution
To get ahead of this problem in 2023, I partnered with Localization to come up with a content strategy plan that determined what content could use machine translation, and what should continue to use human translation.
I was able to get buy-in from leadership.
Impact
Leaning on machine translation saved over 100 hours in writing and editing, significantly saving time, money, and resources. Localization was laser-focused on the content they did need to translate.
Principals for using AI in writing
To help determine what should be localized by a human versus machine, I developed these principals for using AI, which were later surfaced during the Google for Developers rebrand.
AI is best used as a co-pilot, not a replacement.
We should use it as a creative companion to help speed up work and free up time.
Humans still need to edit content.
Humans must fact check, confirm there are no hallucinations present, and ensure that content maps back to priority initiatives and messaging.
Tier 1 content should be researched, written, and edited by a human. However, AI is ok to use during an ideation phase.
Tier 1 content can be defined per project, but most often is marketing landing pages, important messaging, scripts, and anything that’s considered highly visible.
Tier 2 content may be prompt-written with AI, as long as it’s edited by a human.
Tier 2 content can be defined per project, but most often is micro copy, SEO copy, and shorthand product descriptions that appear on various surfaces of a website.
Examples of machine localized content
On less prominent areas such as presenter bios, we felt comfortable designating them as Tier 2 content which could be localized using machine translation. In any given I/O there are hundreds of pages of content and just as many presenters. By leveraging machine translation, we cut down a significant amount of time and resources. Even when a human reviews the content, it’s still faster than translating. (Note: Filmed content is never translated and captions are always in English only.)
