Tilde releases advanced Dynamic Learning feature for optimizing MT performance

Machine Translation
Tilde Machine Translation Dynamic Learning

The latest release of the Tilde MT platform includes an advanced new Dynamic Learning feature, which significantly improves MT performance for translators and post-editors.

Dynamic Learning enables MT engines to continuously adapt through use, fine-tuning a MT engine for particular types of content and improving translation quality over time. This powerful feature works by immediately returning a translator's post-edits in the CAT tool environment back to the MT engine, where they are immediately applied to further translations.

Fundamentally, the feature enables an MT engine to "learn" from each of a translator's post-edits, so that a translator doesn't have to make the same correction twice.

To help translators work more productively, the advanced new Dynamic Learning feature now enables MT engines to provide smarter, more accurate and context-specific suggestions based on previous post-edits, taking into account linguistic elements like word inflections, conjugations, and morphology. Post-edits are now used to adapt and improve MT engines at greater speeds than ever before. 

In an initial study of the effects of Dynamic Learning on MT system quality, conducted on a large translation project of more than 30 000 segments, Tilde’s MT team found that Dynamic Learning could improve MT quality by up to 12 BLEU points – a huge leap in performance quality. Therefore the feature can lead to significant gains in productivity for professional translators.

The feature works best for translators who frequently translate domain-specific texts, as the MT engines can correct frequent mistakes and learn the precise usage of items such as domain-specific terminology, technical jargon, and specific vocabulary.

The advanced Dynamic Learning feature is available to post-editors via the Tilde MT plugins for MateCat and SDL Trados Studio.

Contact us for a demo or to get a free trial