State of the Art Neural Machine Translation

Machine Translation (MT) is the translation of text by a computer with no human involvement. Utilizing neural network approaches to MT, Neural MT (NMT) offers significant translation quality improvement over Statistical MT (SMT). For example, with SDL NMT, we observe impressive improvements on English-German, a language pair that was historically challenging for the SMT technology.

Neural MT - DE>EN

NMT systems also utilize machine learning approaches, but these systems learn higher level concepts for producing translation. The system that is created is a multi-layered neural network that produces translation in a similar processing pathway that the brain would follow. Unlike previous SMT models which are mainly limited to surface text, SDL NMT uses a deep learning architecture capable of learning the meaning of the text which enable the machine to perform the translation task at a semantic level leading to fluent and naturally sounding translation output.
SDL NMT captures also both local and global dependencies and can handle long-range word reordering leading to better translation quality between languages of different word order. See Examples below:
Here, the SMT contains a grammatical flaw, where "eine vielversprechende" (a promising) doesn't agree with the number and gender of "Jahr" (year). SDL NMT doesn't have this problem, because it considers the sentence as a whole and captures both local and global dependencies. Additionally, SMT literally translates “Have” as "Haben" even though this German idiomatic expression doesn't use the word "Haben" at all. SDL NMT doesn't try to translate each English word literally and correctly avoids translating the word "Have" completely.

Here the meaning of the Japanese sentence is not preserved in the SMT translation, confusing which entity is in a state of civil war: it's Syria, not the UNHCR. Also, a translation for the main verb of the sentence 発表した (happyou shita) meaning "announced" is completely left out of the SMT, because the verb comes at the end of the sentence in Japanese and needs to be moved all the way back to right after the subject of the sentence (the UNHCR) in English. SDL NMT can handle long-range word reordering and grammatical complexity leading to a correct translation that keeps the meaning of the Japanese sentence.

Today, SDL NMT is available with SDL Enterprise Translation Server for secured translation needs. SDL also offers customized Neural MT engines that are adapted to the customer-provided data and domains for additional translation quality improvement.