2017/03/30

Neural Machine Translation and Patent Translation - English translation of JTF Journal #288

Masa, Kajiki.  “Neural Machine Translation and Patent Translation”.  JTF Journal.   #288. 2017.  p.14-15.  http://journal.jtf.jp/files/user/JTFjournal288_2017_March.pdf

Neural Machine Translation and Patent Translation

Our Approach before Neural Machine Translation

To be able to provide high quality translation at low cost, technology, collaboration, automation, and innovation is our company’s four pillars of management.  Among them, we actively make use of IT systems.  In order to prevent ‘variation by people’ while shortening the delivery date and cutting cost, we have implemented cloud-type translation systems (i.e. Memsource and SmartCAT), as well as putting more effort into human translation.


Memsource and SmartCAT are both cloud-type systems, which allow machine translation (MT below).  These systems are focused on the Japanese market such as creating Japanese version of their websites.  Latter system in particular, provides free for use only (trial usage only at the time of writing).  Both systems matched with our management policies.
We have been considering implementing MT to our service for the past few years (after some research, I found out that my first participation in a seminar on MT was back in 2012), but the conventional MT did not even come close to a rough translation, in which case we could not make use of.  However, with Neural Machine Translation (NMT below), we determined that a rough translation to a minimum level was possible, and demonstrated the verification trials.

Patent Translation and Neural Machine Translation

In the field of patent translation, proofreading becomes more crucial than the translation.  We always care about how much time we spend on proofreading, such as considering technical expressions and legal knowledge, changing the style to meet the demands of customers, and paying close attention to translating terminology and neologisms related to claims and patents.
By using NMT on patent translation where being “faithful to the original text” is desirable, the work load decreased and the work efficiency has definitely been increasing.  Previous MT required a lot of the post edit (PE) process even in the rough translation level, but with the NMT, re-writing work has steadily decreased.  Regarding technical expressions, terms and neologisms unique to patents, if we prepare macro and programs based on the previous works and glossary, further automation will be expected.  If NMT function improves in the future, it is possible to cut the time and cost spent on the translation itself even more, and the problems we face today will most likely be solved by artificial intelligence.  By effectively using the NMT, more time can be spent on proofreading, and we anticipate that it will also help improve the quality of the final products.
However, without PE, patent translation with sufficient quality is not possible, and it is up to the translator’s hand whether or not it can be used as a patent document for business document, whether it can be translated into appropriate claims for each patent, and whether it can be proofread to be tailored to the customer’s requested style.  If the accuracy of the NMT improves, it will not be a good translation unless we improve human ability to translate and the quality of PE.  We hope to improve both NMT quality and PE by human hand to provide fast and high quality patent translations.
In our company, we decided that NMT accuracy will keep improving, therefore we made a snap decision and proposed a new service.  It is “hybrid translation”, combining NMT and cloud-type translation system with PE by human hand.  Rather than completing the translation with just NMT, we are certain that it is up to PE by human hand to complete the product that meets the customer’s request.
We have received a feedback from a company representative that has participated in our trial.  From the feedback, we have found that if we figure out the “habit” of NMT and rewrite the original text so that NMT can recognize sentences easier, the accuracy of the NMT further increases.  Just like video games, “strategy” exists within the system.  Besides PE, human value exists in such ways.

In the Era of NMT, Human Hand (PE) Becomes Most Important

With the accelerating speed of international business development, “speed and quality” is in demand in the translation world as well.  If highly accurate NMT becomes possible, even shorter delivery date and reduction of cost will be demanded.  Perhaps a case where a translation company is not needed will arise.
However in business texts or in useful patent translation where accuracy is demanded, NMT alone is not enough.  It is essential to PE the sentences translated by NMT with human hand.
Post-editing position specializing in PE has not been established yet, but we think the job will be required in the translation industry in the near future.  It might be a good idea to consider switching your position, or get the training you need to be a post editor while you are younger.  
If you prepare now, you might be able to gain the status of being the first post editor in five years or so.  My emphasis is that having a personal branding may even be possible. 
We are actively implementing NMT and a situation in need of PE is arising.  By creating an environment within the company that works toward the new era, a base to educate our in-house translators to become post editors will be created.  I also believe that by seeing new staff working on PE, we can create an atmosphere to encourage the experienced translators who are hesitant about becoming post editors.  
As of now, NMT itself has not contributed much to shortening of time, and if we reduce the price to half at this moment, we would not be making profits (however, we are optimistic about cutting time by repeated practice), but we are certain that higher goals will be achieved eventually.  By doing so, NMT may hold value as the “future of translation company”, and I believe that our customers will have a new interest in the idea.

Innovation Does Not Come from Same Industries, But From Other Industries

I usually use a watch with life log function made by Garmin at the sports gym or the fitness club.  However, Garmin is originally a GPS equipment manufacturer, not a watch company.  Google providing NMT is a search engine company and not a translation company.  
NMT was also created from industries other than translation industry.  Where and when an innovation is born is unknown.  Whether or not the innovation will be useful in the future is unknown as well, but because we are a small translation company, we would like to actively incorporate the innovations brought on by other industries.  
It takes a long time for a large translation firm to fully incorporate new ideas, and if a small company is waiting around for the large company to make their move, they lose their distinguishing features.  I am hoping for a highly diverse translation industry/community, so although it is still time consuming, we want to continue incorporating MT.

The Future of the Patent Translation Is Hybrid-type Translation, Integrating Humans with Machine

In order to provide low cost and high quality patent translation while helping with the customer’s business expansion, it is crucial to smoothly incorporate NMT and to train the staff to become excellent post editors.  Simple translation is useless as a patent translation, but if too much time is spent on the translation, the customer might lose their business chance.  
NMT is only one of many tools used in the translation process.  We will continue to strive for excellent patent translations through technology, collaboration, automation, and innovation.

Originally written by Masa Kajiki, Founder CEO, MK Translation Firm

Translated by Hiroko Matsuda, in-house translator (JP to EN), MK Translation Firm

0 件のコメント:

コメントを投稿