Machine Translation (MT)
Machine translation was developed for the purpose of automatic translation without human resources. Neural machine translation is a method of machine translation.
Machine Translation Type
There are three main types of machine translation.
Rule-based machine translation (RBMT): It uses grammar rules and dictionary data, which are created using human resources, to create a translation. It requires more human resources than other types.
Statistical machine translation (SMT): A statistical model based on corpus (text data) creates translation. It could be a less accurate translation for language pairs that have different grammatical structures.
Neural machine translation (NMT): A translation is predicted by a neural network created by deep learning based on the corpus (text data).
There is also a hybrid machine translation (HMT) that uses RBMT and SMT. Currently, NMT is the most popularly used.
Neural Network
A neural network is a collection of several units that create neural processes similar to those of the human brain on a computer.
Deep Learning
Deep learning, one of the methods of machine learning, uses a multi-layer neural network. Analysis using deep learning technology is used in various fields, including image recognition, speech recognition, and machine translation.
NMT usage
Google has developed Google Neural Machine Translation (GNMT) based on NMT. Google uses GNMT in Google Translate and publishes an API that allows people to use Google Translate. Microsoft has developed the Microsoft Translator Hub to allow users to customize and use NMT models. SYSTRAN also provides a similar service, SYSTRAN Model Studio.
If you want to develop your own NMT system, you can use TensorFlow, an open-source machine learning library developed by Google. Reference
Training of NMT Engine
Generally, steps are similar in Microsoft Translator Hub and SYSTRAN Model Studio.
- Prepare corpus that matches the translation genre and language
- Clean corpus data
- Train NMT engine with corpus
- Evaluate NMT engine quality by BLEU score (values for translation quality) and human review
NMT quality is highly dependent on the quantity and quality of the corpus, the number of training sessions, and training.