Is Neural Machine Translation the Future of Translation Services?
Neural Machine Translation (NMT) is a paradigm that is going to change machine translation and shape the future of the translation industry. Despite the impressive results NMT shows, it still has constraints preventing it to be a viable option for the translation of all types of content.
Neural Machine Translation—State-Of-The-Art Service or a Luxury?
Neural Machine Translation is based on the method of Deep Learning and is a new type of corpus-based machine translation. Unlike traditional MT, NMT can translate complete sentences and not simply words within them due to the neural networks it is based upon.
While NMT is designed to mimic the neurons of the human brain, in his study Making sense of neural machine translation, Mikel L. Forcada warns that it only vaguely resembles the way people’s or in particular translator’s brains work.
Still, NMT is using neural networks that help predict the sequence of words and the likelihood of its appearance. Initially, the process required significant resources and time to complete the learning cycle and despite that Slator reports that it can produce a text with fewer translation errors and reduce post-editing by 26 percent, it was not a spoon for a small translation company.
Some of the bigger corporations like Google, Microsoft, and Yandex immediately took advantage of the revolutionary process. Google introduced Google Neural Machine Translation (GNMT) in November 2016. Their research on Bridging the Gap between Human and Machine Translation revealed that it can reduce translation errors by an average of 60 percent compared to Google’s phrase-based production system.
While at present most of the MT systems use NMT, the approach is still undergoing development and is not yet ready to be offered as a state-of-the-art professional translation service. The good news is that technological advancement will ease its adoption by the translation industry and NMT will be more of a regular option than a luxury.
Is Neural Machine Translation Suitable for any Content?
A Case Study on 30 Translation Directions carried out by the Adam Mickiewicz University in Poznan and the University of Edinburgh revealed that NMT equals or surpasses the quality of translation provided by phrase-based statistical MT.
The facts of the research, however, reveal that neural machine translation works better with certain language pairs like Chinese and English or English and Arabic languages and not that good with other languages. Further research also reveals that the quality of the translated text depends on content.
Based on some specifically targeted case studies published in A Report from the Frontline of NMT in Multilingual’s January 2018 issue, it can be concluded that NMT can be used for translation only of titles or complete articles, based on the content, topic, and language pair.
NMT still needs improvement in order to stop producing unusable translated segments and therefore be ready to be applied to all types of content.
Where Neural Machine Translation Can Fail You?
As a startup with limited resources for localization and translation, you may want to explore the possibilities of NMT as an option to your international market growth.
While machine translation in general, and neural machine translation in particular, can be an alternative to human translation, there are certain fields where NMT can fail you drastically.
Delip Rao, an AI and Natural Language Processing (NLP) researcher, identifies six areas of concern in his article The Real Problems with Neural Machine Translation:
- NMT is bad with out-of-domain data. In other words, if the system is trained on law, it won’t produce good results in a medical context, for example.
- NMT can’t work well with small datasets. The secret behind neural machine translation is that it generalizes better and produces better results with the increase of input data.
- NMT performs poorly with rare words, which can pose a problem with languages using extensive inflection.
- NMT faces problems with long sentences, which can be a problem for certain types of translation, like legal translation, where long sentences are a norm.
- Alignments pose an issue for the NMT system and hence a verb in the target language may be incorrectly linked to the subject and object in addition to the verb in the source language.
- It is difficult to control quality as beam search used to trim beam width and control the number of choices for the translation of a word seems to have no effect in an NMT system.
Still, neural machine translation can be used as a primary source of translation followed by human editing and proofreading. A trained translator can easily identify the bad segments and simply ignore them and translate the text as they would normally do.
Such a blended approach can be quite useful for startups looking for a more affordable translation model. You can learn more about How Translation Technology Can Help You Scale Your Startup Internationally by clicking on the link and downloading our free whitepaper.