The challenge of machine translation
The ambiguity and variation which are inherent to human language make Machine Translation (MT) an extremely challenging task. Anyone who has ever translated a text will know that the correct translation of a given word or expression can rarely be predicted by simply consulting a dictionary. Translation choices are governed by a wide variety of factors which include not only the context of the original text, conventions in the target language which may be specific to the type of text or its genre, but also extra-linguistic factors, such as the target audience, the purpose of the translation, etc. Given the complexity of the task, until recently, even the best MT technology could not begin to compare with the quality of human translation.
What is neural machine translation?
A new approach called neural machine translation has emerged in recent years, and it may hold the promise of achieving near-human quality in translation. Neural machine translation takes advantage of a new generation of machine learning systems called deep neural networks. These systems use complex internal representations to learn to perform tasks by extracting regularities from data. In the case of translation, the data consists of large amounts of both original and translated texts, and the task at hand is to predict the translation based on the original sentence. Interestingly, the translation process in neural systems is similar to the human translation process. The original sentence is first processed and then transformed into a language-independent representation as learnt by the system during training. From this representation the translated sentence is then generated taking into account not only the content of the original, but also the regularities and conventions of the target language.
Neural machine translation at AT Language Solutions
At AT Language Solutions we are working to provide the best machine translation solutions for our clients. Neural networks present some key advantages when compared to previous approaches to machine translation. As opposed to the previous generation of MT systems that proceed word by word or phrase by phrase, neural MT takes the context provided by the entire sentence into account when translating each of the source words. Furthermore, the abstract intermediate representation of the words allows the system to generalise and correctly translate combinations of words that it has not seen before. Finally, neural systems learn domain-specific behaviour reflecting the subtleties of language use which would be otherwise impossible to capture.
AT Language Solutions has more than 20 years of experience in machine translation services (MT). As a result, the company possesses rich bilingual dictionaries, linguistic information and domain-specific translation data. We combine the latest developments in data-driven MT technology with these invaluable resources, allowing neural MT to work in practice providing the best possible translation solutions which are customised for the needs of the client. Come and have a look at our website and discover our translation services!
Head of MT