Topshop
How we helped Topshop permanently reduce costs while maintaining quality and consistency with adaptive neural machine translation.
The situation
Topshop is a leading global fashion retailer with over 500 shops in 58 countries and a large eCommerce presence. We have has worked with Topshop for over 6 years providing high-quality human translations of brand messages, features, articles, blog posts and product descriptions. The translations needed to be of the highest possible quality to boost conversion rates.
We have become an integral partner in the brand’s aggressive global expansion strategy by providing linguistic support and developing technology to integrate Topshop’s eCommerce platform with our own translation workflow management platform, STREAM.
This has allowed Topshop to increase efficiency, reduce time to market and take advantage of the benefits of translation memory and automated delivery and retrieval of localised content.
With a huge catalogue and new products being introduced on a daily basis, Topshop has a significant ongoing requirement for translations. With recent advances in adaptive neural machine translation, we were able to devise and implement a solution to help Topshop continue increasing the frequency of product updates while maintaining quality and consistency at a much lower cost.
The solution
The team built a dedicated, customised, neural machine translation model – configured for the specific purpose of translating Topshop’s product descriptions into French and German.
A very large dataset of content that had been previously translated, reviewed and rated as being of excellent quality was then used to build an artificial neural network and prepare the system for accepting new source language content and generating translations.
As new content was produced by Topshop, this was translated by the model and the output reviewed by the same linguists and reviewers that have worked on the project and were familiar with the brand’s product range and tone of voice. The linguists rated the translations using a range of criteria including fluency, accuracy, comprehensibility, grammatical correctness and the ability to recognise field-specific terminology, among others.
Any amendments to the machine translations by the linguists were fed back into the system, allowing it to enter a continuous cycle of learning and improvement in order to deliver higher quality translations with every iteration.
The results
After a few months of testing, collecting feedback and further training, the neural machine translation model was found to deliver translations of product descriptions that were very close to human translation in terms of quality, and after editing these were indistinguishable from human translation. The model was subsequently placed into production and Topshop experienced a seamless transition from the old process to the new one.
The neural translation model is now used as a tool to prepare the content for editing and in many cases, the linguists simply have to read the content to ensure it is accurate and adheres to Topshop’s unique style and tone of voice – reducing the time taken to translate the content from hours to minutes in some cases.
