PT EN

How About Working as Machine Coach?

Marcos Chiquetto out 29 2024

Consider this case: an artificial intelligence engine has to translate high-school chemistry questions, such as these ones:

——–

On a given diet yielding 3000 Kcal per day, a 70 kg man excretes 27.0g of urea daily. What percentage of his daily energy requirement is met by protein? Assume that gram of protein yields 4 Kcal and 0.16g of nitrogen as urea.

——–

Calculate the energy of a hydrogen atom in its ground state (n = 1) using the mass of the electron rather than the reduced mass of the hydrogen atom.

——–

Of course, any AI engine will be able to provide a translation. However, there is a critical issue: will it be a good translation? How can we be sure a software tool will be able to provide a reliable translation for these school questions?

I am participating in a project to improve the translation provided by an AI tool. Our translation agency, based in Brazil, has just been hired by a large global multilingual translation company to work in an artificial intelligence development project. The task is to review AI-generated translation of school questions with the sole aim of inputting the human-reviewed translation back to the AI engine, in a kind of training process. Hundreds of school questions, including the two examples above, were translated by the AI engine into several languages and our company was hired to review the translation to Spanish. Of course, the basic assumption is that this feedback will improve the AI translation for this kind of material.

— What is the volume of the project?

According to the client, for school questions only,  they had over 1 million words translated by the AI engine to be reviewed, covering the various school domains, such as Math, Philosophy, Physics, Psychology, etc. 

— One million words? I am not familiar with these volumes. What does it means in terms of pages?

Well, a typical book page contains, say, 400 words. Based on this figure, the material to be reviewed would comprise somewhere around 2500 pages. Just for comparison purpose, an edition of Hamlet, by Shakespeare (Simon & Brown  2011) has  330 pages.

That means the volume of material for reviewing was higher than 7 Sheakespeare books. Only for training the AI in the specific area of school questions!

Of course, this training had to be performed for dozens of target languages. Not to mention the same kind of training for other source languages, such as French, German, Chinese, Spanish, Russian, etc.

And then the same training has to be performed in other areas, such as Business, Medicine, Engineering, etc.

A massive investment!

Wether or not this investment will bring profits, and who will actually earn these profits, is something nobody knows at this point.

Anyway, it is giving rise to a new kind of work for translators: machine coaching!

Well, let’s do it.

Follow us here to see all of our weekly posts:

Talk to us

CONTACT

Rua José Jannarelli, 75, 401
05615-000 | São Paulo – SP
+ 55 (11) 3721-1280
        + 55 (21) 9 8123 1484
info@latinlanguages.com.br
Latin Languages