In January 2018, IBM has launched their Neural Machine Translation engine in a preview mode. We have evaluated it on the 15 of their language pairs we have in our MT benchmark and found 5–15% improvement of the MT quality, which puts IBM to the top tier of Cloud general-purpose MT engines. For three language pairs (en-it, fr-es, and en-pt) they become the most performant engine. Also, IBM now works for Chinese (although on our datasets GTCom and Baidu perform better). However, it’s still one of the most expensive Cloud MT out there with $21.4 per 1 million symbols.
As a sidenote, it’s just a preview, meaning the quality may change after it went to production: it may become better (because the models may be improved) or worse (because the models may be optimized for hardware costs). Let’s wait and see!
