An introduction to machine translation for localisation

gamesindustry.biz:

Kris Trusava

Friday 28th January 2022

Machine learning has made its way into nearly every industry, and game localization is no exception. Software providers claim that their machine translation products mark a new era in localization, but gamers are often left wishing that game publishers would pay more attention to detail.

As a professional localization company that currently is working with machine translation post-editing, Alconost could not pass up the topic. In this article we aim to find out what's hot (and what's not) about machine translation (MT) and how to get the most out of it without sacrificing quality.

When machine learning was introduced to localization, it was seen as a great asset, and for quite a while localization companies worked using the PEMT approach. PEMT stands for post-edited machine translation: it means that after a machine translates your text, translators go through it and edit it. The main problem with PEMT is that the machine translates without comparing the text to previous or current translations and a glossary -- it just translates as it "sees" it. So naturally this method results in numerous mistakes, creating a need for manual editing.

As time passed and technology advanced, NMT (neural machine translation) came into play. This proved a much more reliable and robust solution. NMT uses neural networks and deep learning to not just translate the text but actually learn the terminology and its specifics. This makes NMT much more accurate than PEMT and, with sufficient learning, delivers high-quality results much faster than any manual translation.

It's no surprise that there are dozens of ready-made NMT solutions on the market. These can be divided into two main categories: stock and custom