Whether you want to translate an entire book or just a few articles on your website, the only way to achieve 100% human-readable text is by having a human translator do the job for you.
Most of the time, that means you must pay someone on a per-word basis. For example, if you want to translate English to German, you’d be paying anywhere between 12¢/word and 32¢/word. Unless you are translating a text-light children’s book, your total cost may become prohibitive.
If this was 2010, the discussion would end here.
Welcome to the age of AI and Machine learning. In 2020, a machine can still not translate as well as a human can; however, a machine may just be able to translate well enough to make it cost-efficient to integrate it into the translation process.
The cheapest way to produce a translation is by putting both machines and humans at the task.
First, you feed the source text to a machine or API (application programming interface). The machine returns the machine-translated text.
Then, you give both the source text and the translation to a human translator. Their job will be to “humanize” and edit the text so that it reads as if a human had written it.
For this process to be cost-efficient, the machine must do a good job in the first place. Otherwise, the human translator may spend more time editing than they would spend translating.
This article will answer both questions.
To test the APIs, I took two small paragraphs from a book (with permission from the publisher). It is important that this text, and its translation, is not available on the internet. Google trains its translation machine learning model with publicly available translations. When testing a model, the test sample must be outside the training sample.
You will find images of the German-to-English translations of the text below.
It is fair to conclude that Google’s translation API did a better job than Microsoft, Watson, and Yandex. Or, in other words, our human editor would have less work to do given the Google-generated translation.
It is also fair to say, that it would take a translator less time to edit the Google translation than write a translation from scratch.
Note that I have only tried German to English translation. Different languages are likely to yield different results.
Fab is your man.Tell me more