Translation has been a key feature of marketplaces for millennia. We can’t buy or sell if we can’t understand each other.
Online marketplaces have the same problem as the ancient ones, and at a much greater scale. By 2020, 70% of the 4.5 billion internet users could not speak English, and that number is only growing as the next few billion users join us online.
To translate millions of items into more and more languages, marketplaces like eBay, Amazon, Google Play, AirBnb, Booking.com, Alibaba and Facebook started using machine translation.
Marketplace content can be noisy - it’s often user-generated or in mixed language. And it's very dynamic - for example, a creative name for a technology or product that did not exist this morning. And of course, marketplace content can be domain- or locale-specific.
Luckily, marketplaces have key advantages for using machine translation: there is lots of open data on the web, and they are translating in context - the item description can be very useful for translating the title or review, and the seller’s and buyer’s location and display language are known.
The most impactful translation quality issues in marketplaces are very objective, for example whether Apple is correctly translated as the brand, or incorrectly translated as a fruit.
Machine translation is great when it works. But a bad translation can be worse than no translation.
A bad translation of the product brand name, Self Portrait. What’s the impression on the buyer? What’s the financial and reputational impact on the brand?
Unlike many platforms using machine translation, marketplaces can measure the impact of translation and translation quality on their key metrics, like conversion rates and total sales.
An 2019 study on translation quality and sales on eBay - $14B total - found that just improving translation quality boosted sales between the US and the target markets by more than 10%.
Human translation has the best quality, but it doesn’t scale - it’s slow and 1000x more expensive than machine translation. So how can translation coverage and quality be balanced efficiently?
What if marketplaces could catch and fix the highest-impact machine translation errors?
ModelFront’s translation risk prediction is an instant segment-level signal for implementing hybrid translation - fixing only the riskiest translations or only those translations above a certain risk threshold.
Traditional localization is default-human, and uses machine translation as a minor efficiency. Marketplaces can implement default-machine translation and dial up or down the risk threshold to balance machine efficiency and human quality.
The best strategy varies by the content type, locale and quality goal. Fixing the source once is often more efficient than fixing a bad translation in dozens of languages. Should a bad translation be up until it’s fixed, or is the original untranslated source segment - typically English - a better placeholder? Should a poorly translated review be fixed or just hidden?
What’s the right risk threshold and strategy?
Top-selling items or top reviews have have more visibility. The title has more impact than the description. And descriptions and reviews vary greatly in length.
The value of translation also varies by language and locale. Languages and locales have different numbers of speakers. People in different locales have different expectations of translation quality, and can be comfortable operating in a large lingua franca like English, Hindi, Russian or French.
Advanced marketplaces should have and improve a formula for setting the right risk threshold for every translation. The goal is to maximize the quality experienced across all users and the conversion rate per dollar invested in human translation and post-editing.
Marketplaces are more than product titles and descriptions.
Search, recommendations, reviews, messages between buyers and sellers, email notifications, filtering spam and restricted content, customer service and investigation of fraud and abuse all involve text in many languages.
ModelFront freely shares guidance on how to make these systems and processes multilingual, and will be releasing more open playbooks for them soon. You can read our playbook on multilingual search now.
eBay, as an international two-sided marketplace, was the first giant tech platform to invest in catching bad machine translations, and measuring the impact. Now Amazon and Facebook have also started using that approach.
ModelFront is the first and only provider of translation risk prediction technology, making a production-strength system accessible and useful to more players.