Machine translation has made steady advances over the last few years, and is widely used by professional translators when incorporated into computer-aided translation tools. In recent years the application of artificial neural networks to machine translation has enabled substantial improvements in its efficiency. In neural machine translation (NMT), a large neural network is trained to improve translation performance for particular areas or applications, and with practice it can use its own judgement and become an expert in a specific domain. In effect, the system translates whole sentences rather than individual chunks or phrases.
Google launched Google Translate ten years ago with phrase-based machine translation (PBMT), but recently introduced its own neural machine translation system (GNMT), which is an improvement on PBMT and is now incorporated into Google Translate. Microsoft has also introduced neural machine translation into its Microsoft Translator. This uses a supercomputer for speech and text translation in applications such as Skype, and the quality of Microsoft’s new neural network models can be tested at https://translator.microsoft.com/neural. A third option is provided by the developers of Linguee, who recently introduced DeepL, their version of neural translation. Other companies currently involved in neural machine translation include IBM, SAP, Systran. Facebook, Amazon, Salesforce, Baidu, Lilt and TransPerfect.
Researchers from Google summarised the technical developments in an online article, and summarised the improvement provided by GNMT as follows: „Using a human side-by-side evaluation on a set of isolated simple sentences, it reduces translation errors by an average of 60% compared to Google’s phrase-based production system.“ This is illustrated by the following example:
German sentence: Probleme kann man niemals mit derselben Denkweise lösen, durch die sie entstanden sind.
English PBMT (old): No problem can be solved from the same consciousness that they have arisen.
English GNMT (new): Problems can never be solved with the same way of thinking that caused them.
These developments will clearly impact the future of human translation. According to the Slator Neural Machine Translation Report 2018, „As neural machine translation ripples through the language services supply chain, increasing adoption will disrupt the way the industry operates. … Human and machine interaction is expected to change both to adapt to new neural network-based technology and to improve overall usability of existing tools“.
An unexpected property of single multilingual NMT models is that they can learn to translate between language pairs that they have never encountered. Johnson et al. report that „a multilingual NMT model trained with Portugese > English and English > Spanish examples can generate reasonable translations for Portuguese > Spanish although it has not seen any data for that language pair.“ They refer to this process as implicitly-learned bridging or zero-shot translation.
For an extensive description of the development and implications of Google’s GNMT, see ‚The Great A.I. Awakening‚ by Gideon Lewis-Kraus in the New York Times Magazine of 14 December 2016.
‚Pure Neural™ Machine Translation (PNMT™) Beta Test‚,
‚Round-trip translation: no more entertainment with PNMT™ systems‚
‚A Closer Look at SDL’s new MT announcements‚
‚The Ongoing Neural Machine Translation Momentum‚