Tilde MT Dynamic Learning: Latest Innovation in Neural Machine Translation

Tilde MT Dynamic Learning launches

The European language technology innovator Tilde launches a revolutionary Dynamic Learning technology for machine translation. Tilde MT Dynamic Learning picks up human corrections in real time, thus significantly boosting the quality of machine translation. This novel technology was made possible thanks to the break-through research of the award-winning Tilde research team.

Dynamic Learning leverages Neural MT with translation-memory-like technology. However, unlike translation memory, which only provides translation suggestions from the previously translated text, dynamic learning NMT engines examine the full context of the entire text to provide smarter, more accurate and context-specific suggestions based on previous edits.

When machine translation emerged as a powerful tool for translators, the quality gap between the general and custom MT systems was a notable problem for language service providers. Often, generic systems did not provide the degree of accuracy or terminology required for domain or client specific translations. And developing custom MT engines for each customer and domain was not always feasible or time efficient solution. Tilde MT Dynamic Learning solves this problem and gives translators and LSPs a solution for fast on-the-fly engine adaptation to specific translation projects or domains. It analyzes every correction made by translators and instantly adapts the MT model to meet lexical, terminological and stylistic requirements. With this flexibility, a translator can immediately benefit not only from his or her own post-editing efforts, but also from the work of the entire team of a collaborative translation project.

Tilde MT Dynamic Learning keeps adapting as it is being used. It works by instantly feeding user edits from the CAT tool into the MT engine, where they are used to improve further translations. The engine learns from human feedback in real time. As a result, the next time it already knows how to use domain-specific terminology, technical jargon, specific client vocabulary, etc. accurately. It represents a huge productivity gain for translators, who no longer have to repeat the same correction again and again.

Dynamic Learning is the latest achievement of Tilde research team that has demonstrated top results at the prestigious global machine translation competition WMT for three consecutive years.

An initial study on the impact of dynamic learning NMT engines used in real-life translation projects proves a huge improvement in the quality of machine translation – the average increase is 24 BLEU points, while in some cases the quality skyrocketed close to human parity! 

Comparing BLEU scores of baseline MT system with MT system with Dynamic Learning

Further tests of the dynamic learning functionality during an in-depth study of the translation edit rate (TER) showed that adaptive neural MT reduced post-editing efforts on the average by 0.17 TER compared to baseline MT. Thus, it is safe to claim that dynamic learning NMT engines can deliver significant gains in productivity and efficiency to professional translators.

Comparing translation edit rate of Baseline MT System with MT System with Dynamic Learning

Paula Reichenberg, the Director of Hieronymus, shared her insights into the new feature: “Tilde’s Dynamic Learning Technology opens a new dimension in the Neural Machine Translation. Hieronymus Ltd specializes in legal translations, and we need to keep adapting our terminology and style to various legal sub-domains. The Dynamic Learning Technology will further enhance the performance of our custom NMT-engines by allowing us to adapt them in real time to each specific topic. We are thrilled to be one the first customers to benefit from this new feature.” 

Izabela Radomska, the Head of Project Management from translation agency Bireta was one of the first to get a chance to test the new feature: “We were excited to test Tilde’s Dynamic Learning feature for Polish – English translations. We use general engines for many of our projects, and now with Dynamic feature, we will have an engine that adapts to a specific project considering customer’s style and terminology. We’re enthusiastic about the opportunities this technology could bring in the future.”

Tilde MT Dynamic Learning is available to translators via Tilde MT platform for SDL Trados Studio, MateCat and other. Explore Tilde MT Dynamic Learning with a free trial: https://www.tilde.com/products-and-services/machine-translation