IKEA-MT: the challenges, successes, and pitfalls

Here at IKEA, we have a massive need for translation of content designed to reach both external and internal consumers globally: websites, portals, brochures, packaging, and in-store signs to name but a few. We’ve worked hard to maintain a unique IKEA tone-of-voice in all communications — a polite, inclusive voice with a twist of humour that customers immediately recognise. So, when it comes to translating our ever-growing content into 43 languages, we’re faced with a problem: we want to use the latest neural machine translation(NMT) technologies in order to keep up with demand, but we don’t want our special voice to get lost in the mix of all-encompassing models that everyone is using.

We decided to explore the MT landscape, beyond the major players, to see whether we could use some basic building blocks and then incrementally construct neural network models that better served our needs. And so, IKEA-MT was born — a set of NMT models trained mainly on IKEA texts gathered from years of high-quality human translation content.

In this talk, we describe the early developmental stages of IKEA-MT: the challenges, successes, and pitfalls that we encountered in creating a new technology and then introducing it into our translation ecosystem. The responsible team manager, Johan Sporre, and the lead engineer, Brona Nilsson, will both give their perspectives on the journey so far and speculate on what the situation will look like in the future.

About Brona

Systems Engineer, Global Language Services at Ingka Group. Using machine learning to connect people.

Currently, she is training and deploying neural machine translation models to help serve the voracious demand for translation within Ingka group and Inter IKEA.

In the past, she worked for 15 years in speech & language technology companies including Nuance Communications, SVOX, and Lernout & Hauspie. She holds a Ph.D. in AI from Trinity College Dublin.