How much does it cost to translate Habr?

Hi, Habr! The guys from Smartcat decided to go crazy and try to translate into English all the posts that were published here before July 19, 2017, and then estimate how much it will cost on average if the VS Man Machine translates. Under the cut, you will learn what came out of it.



Without unnecessary introductory, I pass on the word scalywhale from Smartcat.

8,729,613 words

Or 62,397,253 characters - so much text is on habrahabr.ru .
')
The most common practice among our clients (mainly translation companies): first, the text is translated, then the editor checks it, and then the corrector rules. Let's leave only the translation stage, because the speed of content delivery is directly proportional to its value, and on Habré hardly all texts go through so many iterations.



2500 words per day - the translator can work with such speed on average, which means that he would have managed to translate the entire Habr for seven years and six months without days off and holidays. During this time, translatable texts will lose relevance, plus new ones will be added to them, and the translator will likely go crazy .

Translates Man

An experienced English translator takes an average of $ 0.08 ($ 4.80 *) per word, which turns out to be $ 698,369 ( $ 41,874,973.45). Excluding the costs of process management.



Let's try as quickly as possible. Let several translators take on the project at once, in Smartcat you can work on one project and even a document together at the same time. We will assemble a team of 50 people, their total productivity will be 125 thousand words per day. This means that the translation itself will take 70 days, and the cost will remain the same. Let's add here two weeks to find suitable candidates and testing, and this is the minimum.

SMT vs NMT

So, let's try to deal with the task even faster and use machine translation. The technology, which, in the opinion of most Internet users, needs only to quickly and literally translate something, has recently begun to translate so well that the translation industry has seriously paid attention to it. One need only recall the news at the end of 2016, when the news first appeared that the translator from Microsoft now works using a neural network , and then an article from The New York Times appeared , in which they told that Google translate learned to translate texts almost like a man.

Previously, machine translation engines used algorithms based on rules and statistical models derived from large volumes of bilingual texts, which is why it is called Statistical Machine Translation (SMT). The new technology uses an artificial neural network, which independently studies the deep connections in languages ​​at the level of whole sentences, rather than individual phrases, and on their basis generates more accurate and readable translations.

Translates Machine

In general, machine translation comes into play. For clarity, let's take this article with Habr of 842 words . A live translator will handle it in about three hours and ask for $ 67.4 ($ 4,041.38). Let's make it easier for him and at the same time save. We will entrust the translation to the machine, and the editing to the person.

This translation method is called post-editing and requires special skills. The post editor should not only speak the language, but also understand how machine translation works.

So, we connect the machine translation and look for the editor through our site Smartcat . Fill in the Word document with the text of the article and tick the pre-translation via Microsoft Translator. On the site, you can not only translate, but also look for freelance translators from around the world, including post-editors with native English.



Services of editors cost less than translation services. We found a freelance girl asking for $ 0.022 ($ 1.32) per word. In the amount of text translation costs $ 18.5 ($ 1109.28). According to the very same editor, she coped with the task faster than if she translated it herself - in 2 hours. As a machine translation engine, we used the paid version of Microsoft Translator , which should translate better. Recalculate the figures at the rate of S1:



As a result, the transfer according to this scheme is 75% more profitable and a third faster. It turns out, if you use machine translation and hire 50 editors, then the whole Habr can be transferred in 48 days for $ 192,675 (11,553,004.94 ₽).

Opinions

Is the use of machine translation in the professional field uniquely effective? We have collected for you a few opinions from our customers.

Aleksey Dyagterev, head of B2B-Center electronic trading platform, says that they are trying to attract foreign companies to the site. Previously, only the texts of the most significant procedures were translated into English by hand, about 10% of all lots. Now, thanks to the machine translation of an international audience, the titles and descriptions of all 5,000 lots published daily on the site are available. The quality of translation is acceptable - it is enough to find out the information and then specify the details.

“ Thanks to machine translation and integration with the Smartcat system, routine operations are performed in an automated mode, and the efficiency of the use of qualified employees has increased significantly .”

Fedor Bezrukov, head of the department of one of the largest Russian translation companies Logrus IT , argues that there is an edge to the new technology, but not everything is so simple.

“ Recently we received an urgent order for the translation of a technical text of 900 words from Russian to English. Three engines of machine translation were connected at once - statistical (SMT) and neural (NMT) from Microsoft and statistical from Google. And for checking stylistics and grammar - also the Grammarly plugin. The Microsoft NMT and Google SMT gave the most successful translations. The translation was ready in 1 hour and 40 minutes, the process was controlled by the translator. It turns out, we have achieved performance ≈ 500 words per hour . "

According to Fedor, the difference between statistical and neural machine translation lies in the fact that neural translators produce a much more coherent text, but they pose a danger: the result may be well-read nonsense.

“At this stage, we prefer to use the issuance of several engines to combine the advantages of each and level the disadvantages. When NMT-engines can be trained and trained terminology on the fly, the process will reach a qualitatively new level. ”

Recently, colleagues from Weebly turned to us, who decided to localize their product in 13 languages. It immediately became clear to us that the project is not only ambitious, but also complex - the text content of the site is distributed throughout the system and stored in various formats, plus it is constantly changing and being updated. An elegant solution was found: thanks to the integration of the Weebly website through the API, texts were easily processed, translated and transferred back to the website. To speed up the work, a team of 5–10 people worked on the translation into each language, which is a good hundred translators. We actively used machine translation - so that the translation went faster and to check texts in different languages ​​on the layout.

“ The Smartcat team supported us at every stage. Whenever there were questions or suddenly new tasks appeared, we could count on the fact that the guys would help or share their experience. Thanks to Smartcat, we were able to localize the Weebly website qualitatively and in a short time into 13 languages, effectively managing the process at every stage: starting with searching translators and distributing tasks and ending with data management and integration of automated solutions into the project . ” Nicolas Olucha Sanchez, localization manager at Weebly.

“The Weebly project turned out to be difficult, and therefore interesting. At Smartcat, we are developing a translation community and creating smart technologies, enabling companies to scale their businesses with ease. With us you can easily find a performer or put together a whole team, combine machine translation engines, connect glossaries and translation memory, and if there’s a lot of work, we’ll do everything for you. We love challenging tasks, if you have any, write! ”Sergey Andreev, product manager at Smartcat

* Throughout the article, the conversion of $ to ₽ at the rate of the Central Bank of the Russian Federation on August 10, 2017. Data from the site .

about the author

Pavel Doronin - loves localization, translations and everything connected with it, and works on creating the best tools for it. He also loves electronic music and synthesizers (after work). # i18n # l10n # xl8n