Machine Translation and Document Localization at Adobe: From Pilot to Production

Adobe recently began exploring the integration of Machine Translation (MT) technology into its localization workflow. The primary question that we sought to answer was whether post-editing Machine Translation output was faster than translating the text from scratch. The exploration occurred in two stages: a small pilot followed by a larger project localizing Adobe product documentation. The project used two MT engines (one statistical and one traditional) both of which were trained with Adobe data and lexicons. Initially, a small test set of 800-2000 words of documentation was machine translated and post-edited, and based on the positive results we proceeded to the second stage, localizing about 200,000 words of new text. The second stage completed successfully, but with some complications, including: • The post-editing rate and the MT quality varied significantly between files. Additionally, differences in the first and second pilot test data created differences in MT quality. • The technical integration of MT with the localization management system has been much more difficult than expected. • The translation vendor reported that postediting required more-highly-skilled translators to perform the task, which was contrary to our expectations.