Translation between different human languages was always important part of many business or other private, social activities, and it is becoming more so in the context of intensifying international interaction. Computational linguists have been experimenting with many different approaches but in the recent decade the breakthrough with the help of computational natural language processing and especially deep learning provide access to such services with reasonable result.
Although all major IT companies provide general solutions -- sometimes even for free --, there are still circumstances where specific product development is needed, e. g. in case of smaller languages (like Hungarian) involved, specific topics (e. g. medicine or law) or sensitive data which mustn’t be leveraged to external providers. In these cases, unique, on premise, tailored solutions could be developed in the specific form of machine learning: neural machine translation, based on the necessary amount of parallel training corpora of source-destination text (usually sentence pairs.
Challenges of Using Neural Machine Translation
The challenges of such a development are the necessary amount and quality of training data (usually on the magnitude of millions or tens of millions), computational power, the difficulty of proper understanding of human languages with all of its ambiguities and the problem of objectively evaluating different translation alternatives (several valid translations could exist for any human expression which are quite difficult to compare qualitatively).
Added values (Why AI/ML/DL):automatization of human language translation for saving cost and time.