[Harsh truth] Introduction of MTPE doesn't make translation speed any faster and it won't be any cheaper.

This is truly an unfortunate situation to write about at the end of 2017.

Introduction of MTPE does not make translation speed any faster or any cheaper. I will say this again.

Introduction of MTPE does not make translation speed any faster. Therefore, the price won't go down. 

This is the conclusion we came to after spending a year in 2017 for verification. We tried to use one project and use NMT to some how make the translation speed faster and the price cheaper. In the end, we decided to step down from MTPE.

Most likely, people in the same business as us (translation companies specializing in patent translation) have already noticed. When NMT was made available, the whole industry became excited. Each company came up with services using NMT in various ways. However, it is clear from the fact that MTPE is not offered as a product in the patent translation industry (perhaps this is good news for translators. This means that translators won't lose their jobs by NMT).

NMT cannot translate text using complex sentence structures seen in patent specifications. At first glance, you might think it is translating, but it doesn't come close to a level of "translating" for us. This is because correction is needed across the entire text.

Of course, it may be different between engine to engine, but translation output from NMT is TOEIC 800 level at most. Even if a translator at this level translates (at least regarding patent specification), we can only expect anything but a proper translation. I would be terrified. I would consider rewriting the whole text.

The biggest reason for improper translation is that translation result cannot be controlled. Even a new translator who can only get TOEIC score of 800 can unify terms when they are instructed face to face to unify "determine" being translated as 「判別する」. NMT cannot do this (Adaptive Machine Translation (AMT) can).

Therefore, there is not much of a difference in speed whether humans translate from scratch or using NMT to perform post editing. When MTPE is utilized, upon verifying within the company, it was confirmed that productivity improved to a degree of 20%. However, if we set the target of productivity improvement to 20% or so, combining the existing voice input software, software such as translation memory, and autosuccession function (?) may be enough to achieve this. There's no reason to even bother introducing NMT.

From 2018, we are considering canceling MTPE service for now. If the situation changes, we might reconsider then.

This past year, we made various efforts to incorporate NMT as one of the services. As a conclusion, we chose to step down from it. Starting from 2018, we will be switching our gears to introduce a new service.

*The above article is written only about translating patent specification. When general text is being translated, NMT is effective to an extent.
*Under a condition in which the original specification text is written with post edit in mind, using NMT might be meaningful. Therefore, it would be more meaningful if the text is written under the environment in which the applicant writes the specification and translate within a patent office.


Above is translation of an article "【残酷な事実】ＭＴＰＥを導入しても翻訳速度は早くはならないし、値段も下がらない。" dated December 30, 2017
Translation by Hiroko Matsuda

What is Adaptive Machine Translation (AMT)?

I'll write about the difference between Neural Machine Translation (NMT) and Adaptive Machine Translation (AMT).

In NMT, when machine translation is executed on a certain text, you cannot control the term translation. Something like this:

English (source)
.......determine............................................................................................... .....................................determines................................................................................determination................................................................................. .....................................................................determined.....................

Japanese (target)
.......判別する............................................................................................... .....................................判断する..................................................................判定する................................................................................. .....................................................................判別する....................

In this manner, there is no uniformity among target translation terms (unable to enforce uniformity).

In AMT:
English (source)
.......determine............................................................................................... .....................................determines................................................................................determination................................................................................. .....................................................................determined.....................

Japanese (target)
.......判別する............................................................................................... .....................................判別する..................................................................判別する................................................................................. .....................................................................判別する....................

When the first "determine" is translated and confirmed with 「判別する」,
the machine translation learns that term and will display the term as「判別する」in the subsequent segments.

In terms of work flow, machine translation is not executed on the entire text in the AMT. The machine translation is executed segment by segment.

Therefore, although the term "post edit" is used on NMT, when AMT is being used, it is an actual "translation" (there is no concept of post editing on the latter).

*We are currently in the process of verifying AMT. If various problems can be solved, we may implement AMT in the first half of 2018.
*Since the name includes "adaptive", I think 「適応性機械翻訳」is correct. However, considering its function and contents,  the name「学習型機械翻訳」（Learning type machine translation) is not wrong.



Above is translation of an article "Adaptive Machine Translation (AMT: 学習型機械翻訳)とは？" dated December 17, 2017
Translation by Hiroko Matsuda