Machine Translation: What is the state of the art?
6. November 2019Machine Translation: What is the state of the art?
A new set of machine translation (MT) tools driven by Artificial Intelligence is already translating tens of millions of messages a day. Proprietary ML (Machine Learning) translation solutions from Google, Microsoft and Amazon are in daily use. Facebook is going its way with open source approaches. What works best when translating software, documentation and natural language content? And where is the automation of AI-controlled neural networks?
William Mamane, Head of Digital Marketing at Tomedes, a professional language service provider, was a skeptic of machine translation. His position today: „Our company has been around for 12 years, with more than 50,000 business customers. We have stood up for the value of „human translation“ and we still do. However, we have seen a steady increase in the quality of machine translation. Currently, machine translation does not compete with a good native linguist.“
This raises the question of the means by which AI translators work. First, at a basic level, MT uses algorithms to replace words in one language with words in another. This turns out to be insufficient. Understanding entire phrases is required for both the source and target language.
Second, it uses statistics to select the best translation for a particular phrase or sentence. Some use structured rules to select the most likely meaning. But in complex language forms like fiction or other types of literature, even the best machine translation machines don’t sound natural.
Machines are good where they encounter structured language. These include weather reports, financial reports, government minutes, legal documents, sports results. Language and idioms are limited in these cases. There are formulaic linguistic structures and formats.
Neural Machine Translation (NMT) uses an artificially generated neural network. This deep learning technique considers whole sentences, not just single words, when translating. Neural networks require a fraction of the memory required for statistical methods. They work much faster.
Deep learning or applications of artificial intelligence for translation first appeared in the 90s in speech recognition. The first scientific paper on the use of neural networks in machine translation appeared in 2014. The article was quickly followed by many advances in this field. In 2015, the first NMT system appeared in Open MT, a competition for machine translation. Since then, competitions have been held almost exclusively with NMT tools.
The latest NMT approaches use a so-called bidirectional, recurring neural network (RNN). Microsoft relies on RNN in Microsoft Translator and Skype Translator. Both aim to realize the long-cherished dream of simultaneous translation that the Harvard NLP Group recently released with OpenNMT, a neural machine translation system using open source technology. Facebook is involved in extensive experiments with open source NMT and learns from the language of its users.
Google Translate is a free machine translation service developed by Google to translate text. It offers a website interface, mobile apps for Android and iOS. Google Translate supports over 100 languages. Although the tool is not as perfect as a professional human translation, it is getting closer and closer. In a 2018 study, Google asked native speakers of each language to rate Google Translate’s translation on a scale of 0 and 6: it averaged 5.43. Performance varies by language. For example, Google Translate was tested for its best performance when English is the target language and the source is European.