Tilde Neural Machine Translation

Neural Machine Translation (MT) is the next generation of machine translation technology. Systems built with neural networks produce more fluent translations than ever before, substantially boosting the level of MT quality and accuracy. Tilde developed the world's first neural MT systems for smaller languages.

WHAT IS NEURAL MT?

Neural MT is based on neural network technology, which is a high-powered approach to Artificial Ingelligence (AI) and machine learning. As opposed to traditional statistical MT methods, neural MT examines an entire sentence before suggesting the best possible translations. This new model provides translations that are more fluent and readable than ever before. Tilde Neural MT achieves this success by taking into account elements like word reordering, dependencies, morpho-syntactic agreements, and other features that are not handled well by statistical MT.

BENEFITS OF NEURAL MT

  • More fluent, human-like translations
  • Higher translation quality and accuracy
  • Context is taken into account
  • Better for under-resourced languages with less data

Try the world's first neural MT systems for smaller languages


COMPARING SMT WITH NEURAL MT RESULTS

To compare the translation quality of traditional statistical machine translation (SMT) engines with neural machine translation (NMT), Tilde enlisted professional translators to perform a comparative evaluation and error analysis of translations. The translators preferred NMT over SMT in multiple language pairs. They also found that NMT systems are more fluent and precise, as well as up to five times better than SMT systems at handling elements like word ordering, morphology, syntax, and word agreements.

Read the results of a comparative evaluation between SMT and NMT

 

Old translation

 

smt_final_1.jpg

 

New translation

 

nmt_final_1.jpg

 

 

 

Old translation

lv_smt_final_1_0.jpg

New translation

lv_nmt_final_1.jpg


TECHNOLOGICAL FOUNDATIONS

Tilde is at the forefront of research and innovation in neural machine translation in Europe. The result of Tilde’s groundbreaking research has been shared with members of the Quality Translation 21 (QT21) project, funded by the European Union’s Horizon 2020 research and innovation programme. Tilde plans to roll out neural MT systems for other smaller languages – considered to be languages with less than 8 million speakers – over the coming year.

Tilde Neural MT technology is based on seamless integration between:

  • the Tilde MT platform, which provides workflows for data pre- and post-processing (cleaning and filtering of training data, language-based tokenisation and de-tokenisation, named entity and non-translatable token management, etc.); formatting tag management, which is a must when dealing with document translation and using MT in professional translation environments; integration in popular CAT environments; and many other useful features; and
  • Neural MT systems with sub-word level attention based encoder- decoder architecture that uses gated recurrent units.