Neural Machine Translation (NMT) is like politics: everyone talks about it, but few know what’s truly going on.
President of Local Concept, Michael R. Cárdenas, and President of Systran, Dennis Gachot, presented at this year’s LocWorld in Kuala Lumpur on the status of NMT. The discussion outlined an objective view on this technology.
What is Neural Machine Translation?
In short, it replaces traditional, statistical MT with a Neural Network model. NMT is known to create more accurate output than Statistical MT, however, it is not for everyone. You have to dive in and get your feet wet before concluding if it will work for you or not.
What is the difference between Neural Machine Translation and Statistical Machine Translation?
While MT uses algorithms purely based on statistical models, NMT learns linguistic patterns and applies them to translate text. In other words, the neural network can be trained to recognize data patterns and improve translation output over time, whereas Statistical MT uses the most probable output.
How do I know if MT is for me?
The standard statistical quality analysis methods, such as BlueScore, are a starting point for quality analysis, but you need to follow up with data analysis from a human-based quality metric.
While it works considerably well for technical text, creative material still sees very weak results. The quality is also different per language pair. To effectively rely on NMT for technical material, there needs to be a substantial investment of time and money to train the engine for your language pair(s).
Is NMT here to stay?
Yes, it definitely is. More and more research is being done each day and the advancements in the area are noticeable. If you want to remain ahead of the game, you need to get your feet wet now.
Any questions about NMT or MT? Leave a comment below!