As you've seen, neural machine translation through Watson is powerful. You're not limited to just word-based or phrase-based translation with Watson. Rather, it tries to understand the true meaning of the input; then, outputs the same intent in another language, even if it has different wording or sentence structuring.
However, natural language is a very broad domain; imagine trying to squeeze the entirety of human language —all of the expressions, vocabulary, domain-specific phrases, and more—into just one dataset and machine learning model. That's, unfortunately, not possible—there are just too many domains and fields, where specific lingo or jargon may be used.
That's why, with Watson, you're not just limited to the pre-trained models—you can train your own models, to tune the language in your own domain!
For example, as you can imagine, the word usage and sentence structure of United Nations speeches are different from the average email. Therefore...