Translated> How many machine translation engines should we use? 20170924

I'm back from the Machine Translation Summit. Looking back at my notes, I wrote down things like these on post editing (well, typed on PC actually).

1. Domain specific machine translation engine is better.
2. It's possible to construct our own engine as long as corpus can be accumulated.
3. Specific operation details are trade secret.

In regards to 1, it's no use having machine translation aimed for general purpose. Accuracy will be higher with machine translation engine of specific field type. 

Regarding 2, corpus does not exist for language pairs other than En/Ja and Ja/En, so we need to construct our own. For latter pair, the server fee is becoming lower, and it's possible for smaller companies like ourselves to possess in-house server. In addition, with Japanese clients, there are many cases that clients cannot request projects if the server is not within the country.

For 3, much consideration is needed. All we need to do is to combine multiple machine translation engines and use the best one. Some CAT tools don't allow multiple APIs to be constructed, so that is one problem to be considered. Whether the server is located domestically or internationally should be publicized, but the combination of engines and how the translation is operated may be considered trade secret.

To be continued...