Machine translation (MT) has been viewed with a critical eye at the beginning of its existence, however, professionals and users of the language industry soon learned how to harness its power properly. With the technological advancement, the new paradigm that has emerged in the translation sector is neural machine translation (NMT). It is a far more advanced tool that can offer great translation results but is still far away from replacing human translation.
Before you decide whether using machine translation and more specific neural machine translation is right for your translation project, it is necessary to know more about it. Note that NMT has evolved significantly after it has been introduced for the first time, however, it still has some shortcomings and if blindly trusted can fail you.
What is neural machine translation?
Neural machine translation is a process that is based on the method of Deep Learning and Representation Learning and is a newer form of machine translation that uses large artificial neural networks to provide a translation. Its main difference from the traditional MT is that it predicts which is the likely sequence of the words and thus translates the whole sentence and not only the words it is comprised of.
The idea of deep learning applications was first mentioned in the 1990s in relation to speech recognition. Later, in 2014 the first paper on machine translation using neural networks was published, which set the beginning of the rapid development in the sphere. In 2015, an NMT system participated for the first time in a public machine competition.
Since the very name of NMT contains a reference to the neurons of the human brain, many people expect or are afraid of the technology replacing human translation. While there is a resemblance in the way the machines work to produce the translation, in his study Making Sense of Neural Machine Translation Mikel L. Forcada stresses out that it is only vaguely mimicking the way the brain of a translator works.
One of the main advantages of NMT compared to the traditional statistical machine translation, is that it uses a much smaller amount of memory and all parts of the model are trained jointly. This results in a better translation performance.
Can neural machine translation be the future of translation services?
With the development of NMT, it started gaining popularity and some of the large corporation such as Google, Microsoft, and Yandex quickly took advantage of the revolutionary model.
Google Neural Machine Translation (GNMT) was introduced in November 2016 and according to research on Bridging the Gap between Human and Machine Translation carried out by the company, the new model can result in 60% fewer translation errors compared to the phrase-based production system Google used prior to that.
Lots of research was carried out in the field to determine how reliable NMT is and whether it can be applied in all types of texts. A Case Study on 30 Translation Directions carried out by the Adam Mickiewicz University in Poznan and the University of Edinburgh revealed that NMT can offer equal and even better translations that the traditional MT used but that is functions better only with certain language pairs, English-Chinese and English-Arabic being one of the most successful. The results also showed that the quality of the translated text also greatly depended on the content.
A Report from the Frontline of NMT in Multilingual’s January 2018 issue also provides the same conclusions. NMT can provide efficient translation but it depends largely on the content, topic, and language pair used.
Having said all that, it is also fair to point out that currently most of the MT systems use NMT, which is undergoing fast changes and improvements. The approach can be quite helpful for the translation industry and enable translation companies to offer more affordable and fast services that can be part of the regular options provided to customers.
Which are the weak points of neural machine translation?
While neural machine translation evolves constantly, it is still far away from the high quality of human translation. There are certain weak points that you need to be aware of, should you decide to consider NMT for your next translation project.
One of the best descriptions of NMT shortcomings is given by Delip Rao, an AI and Natural Language Processing (NLP) researcher, in his article The Real Problems with Neural Machine Translation. He identifies six areas where NMT can fail you, which are as follows:
- NMT performs badly with out-of-domain data.
- NMT needs a larger database to work well and fails to perform well with smaller datasets.
- NMT does not cope well with rare words.
- NMT performs poorly with longer sentences.
- Alignments are an issue for NMT
- Beam search cannot be used to control quality within an NMT system
These are just the main points that you need to keep in mind should you decide to trust your entire translation to a machine.
Normally, translation agencies offer a form of hybrid translations, which involves human translators proofreading and editing the output produced by the MT. In this sense, NMT can provide an output of higher quality that can facilitate the overall process. Bad segments will be easily identified by the trained translator and eliminated.
When NMT is a good option?
Depending on the industry you are in, MT can be a viable option that can save you lots of time and money. The travel and tourism industry is one of the fields, where this approach can be very successfully integrated.
NMT systems can be properly trained and used for processing travelers feedback posting on a tourist agency website, for example.
NMT can be used by industries using a lot of technical documentation to translate documents for internal use.
NMT can be used by any business to get the gist of the new tendencies in the field but should not be trusted blindly as it still can produce some inadequate translation that can be misleading. Hence, if you need document translation services that are reliable and professional, ICD Translation is here to help you in any language you require.
Making Sense of Neural Machine Translation by Mikel L. Forcada
A Report from the Frontline of NMT by John Tinsley, published in Multilingual’s January 2018 issue
The Real Problems with Neural Machine Translation by Delip Rao