In the last issue of our newsletter, we tackled the looming question of AI and its potential to shake up the field of academic publishing. This time, we take a look at machine translation. Tools such as Google Translate have been around for years, and have now been joined by such newcomers as DeepL. The techniques utilized to power these tools have increased in sophistication over the years, with most now using variations of machine learning and neural networks. This is particularly exciting, given that while English is the predominant language of research publications, much research is conducted in non-English-speaking countries. The inequity this has produced is something that some researchers believe machine translation could help ameliorate in the future.
The primary concern with machine translations, however, is accuracy. For example, a great deal of linguistic information is contextual, and much is also often omitted when the writer can assume the reader will draw the correct inferences. These, among other issues, have historically been problems for machine translators and have been discussed broadly.
This might soon change, though, with improvements in machine learning techniques. Anecdotally, platforms such as DeepL can show impressive accuracy compared to long-timers in the industry such as Google Translate (which has itself improved greatly over the years), producing surprisingly idiomatic translations between difficult language pairs such as English and Japanese. Although data is yet sparse, there are some preliminary analyses that shows these rumors and firsthand accounts may hold some water, including one from a Japanese research team.
Despite such exciting developments, it must be noted (as also suggested by the authors of the aforementioned analysis that machine translations still face many limitations, especially when it comes to expert domain knowledge and highly complex sentence structures. While machine translation solutions are tantalizing, it is still necessary, for the time being, to ensure that translations are reviewed by human domain experts.
Click here for the Japanese version.