“Crocodile Skin Case” or “Alligator Bag”: Comparison of Online Translators Connected to Lokalise
Lokalise users can choose whether to localize their product with the help of hired interpreters of the site, with their own team or exclusively on their own. It is to simplify the procedure of localization of those projects where professional translators are not needed and own knowledge of the language is enough, we provide our users with the opportunity to use the popular machine translation systems from Google, Yandex, Microsoft and SDL built into Lokalise. How these systems are translated, today we will talk on concrete examples.
About six months ago, Google announced the connection of another set of languages to the neural network of its Google Translate service, including Russian. This event has become a landmark for the entire Russian-language Internet space: every day thousands of people use Google’s built-in Google translator or go to Google Translate to translate a foreign text into their native language.
Google Translate still retains many features of the classic machine translator based on the rules: when translating a word, the service displays, in addition to the main, most acceptable translation variant of the word, and all its meanings as different parts of speech. In one of our past publications, translations, we have already talked about neural machine translation and how it works.
')
Neural translation, unlike statistical rules based on rigid rules, uses a complex and not always obvious vector system of comparisons in order to choose the most suitable translation based on the context. Of course, it’s not always successful, but more often than not, Google Translate is quite successfully coping with its tasks.
Analysis of the context in the translation field in Google Translate is very easy to check. For the test, we introduce a few words and phrases that can be used in the interface of 90% of applications and see what the translator from the search giant offers:
It seems that the task is extremely simple, but the intellectual side of Google Translate is manifested in one small nuance: the translation of the word “Home”. Based on the context, the translator proposed the option “Home”, which comes closest to the meaning, if we are talking about the software interface in terms of use in the open spaces of the network. At the same time, if we eliminate the context in the form of “app settings” and other words, the translation will remain unchanged - “Home”. All other options in the form of "home" and "dwelling" will be offered in the form of a list under the translation form:
To achieve “machine” from Google translate in the usual sense of the word “Home” translation as “Home” or “Home” is possible only by adding the context - the minimum, in the form of the preposition “at”. With the pretext of Google Translate, it is friendly and instantly translates "at home" as "home."
We will give the translator of the search giant a more difficult task, namely the first sentence of Theodor Dreiser’s novel Sister Kerry. In the original, it looks like this:
In the now classic Soviet translation of Mark Grigorievich Volosov (1895-1941), this sentence looks like this:
Here's how Google Translate handled the translation of this not so simple sentence:
Objectively, the result is excellent for the car. In the preferred translation, the network has lost the word "train", but it is present in one of the options. Also, problems arose with "imitation alligator-skin satchel" and "four dollars in money". The rest of the network performed the translation at a more than decent level for understanding, which requires only stylistic processing.
The Yandex company likes to position itself as the best in everything within the Russian market, and even bypassing the prohibitions, FAS says it is “search number 1”, for example. Whether this statement is true for other products of the company, it will be possible to check at work the built-in translator “Yandex”, which is located even on a domain similar to its elder brother from Google - translate.yandex.ru .
Much has been written about the translator of Yandex, and by the developers themselves. On their own site, the company's developers immediately admit that the translator of Yandex is a statistical one. To make it clearer and not have to go back to the text about the translators mentioned above: in short, statistical translation involves the enumeration of all possible options with the calculation of the most appropriate. You can also bring the whole slide that the team drew to more clearly show the way the Yandex translator works:
This translator initially loses Google Translate, because it does not have a self-learning neural network. However, the Yandex team had at least one trump card: they could try to make the translator’s work on the “English-Russian” pair as correct as possible, since, obviously, the main target audience of the product is Russian-speaking users.
So, let's feed Yandex to the Translator the same set of words with which we initially tested Google Translate:
As a result, we have two misses: "Options" instead of the expected "Options", "Home" instead of "Home". At the same time, the translator persists and offers the translation of “Home” in the context of “Home” only in conjunction with the word “menu”. In other translation variants, the word “Home”, the variant from “Main”, so conveniently slipped into Google Translate based on the needs of the Internet audience, is completely dropped somewhere to the very bottom. The translator offers only the generic version of "House", "House" and so on.
Now let's see how Yandex machine translation will cope with Dreiser’s work:
In the long run in the form of complex constructions, the problems of older technology become more pronounced. Yes, a few years ago, Google Translate translated the same way, but now the difference, as they say, is “obvious”.
The Yandex translator is experiencing the same problems that people who are learning Russian are faced with: it translates too literally, using less context, and does not understand the cases. In general, cases are the pain of any foreigner. We, as native speakers, who have been practicing the proper use of cases since childhood, do not understand how difficult it really is. We think that there is no sense in analyzing all the errors of the Yandex translator: the nature of the machine statistical translation is very clearly shown in the example above. The only thing with which the domestic child did a great job was the form of recording the name of the street.
Yes, the developers say they use both machine learning and dictionaries. However, it is obvious that for processing complex texts or understanding the context, this is not enough.
The next on the list of online translators connected to Lokalise is a product of Microsoft, which is used by their search engine Bing. By the way, the Bing translator is more popular in the world than you can immediately think of: it is embedded in all official applications for Windows Phone (for example, in Twitter), used with the default search in Edge for Windows 10, “hooked” to Skype.
Another distinctive feature of Bing is the use of neural networks. First, the neural networks were hooked up to the Bing search to make it more intelligent and expand the page indexing table, which was two times smaller than Google’s. This happened in early 2015. Later, neural networks reached the Bing translator . Thus, it can be argued that the Bing translator is practically in the same "weight category" with Google Translate.
As before, we first check how Bing Translator will work out a simple set of words of interest:
In general, the translator from Microsoft did better than the system from Yandex, but at the same time he allowed himself a lot more liberties than Google Translate. In the case of our stumbling block - the word "Home" - Bing tried to get out and offered the option "Home", which is associated with the button "Home", but certainly not with "Home" and "Dwelling" in the case of "Yandex". On the other hand, the controversial translation of “Tools” as a “means” raises some questions. At the same time, the Yandex machine translator unequivocally translated it as “Tools”. Perhaps Bing is even too creative, so it’s worth checking out with Dreiser:
In defense of Bing Translator, you can immediately say that so far from all three translators, only he correctly grasped the essence of the phrase “the afternoon train”. The Yandex translator made this afternoon train an evening train, and Google Translate pretended to “pretend to be a hose” and ignored the existence of the train itself. Like his brethren, Bing stumbled over a “cheap imitation bag with an alligator”, and also messed up with “equipment” and “trunk”.
In general, we can say that Bing did a little better than Yandex, but Google Translate translates difficult constructions very far to Google. Probably, there is a lack of sampling and examples for the Bing neural network, since Google translator is obviously much more popular and has great resources.
The latest translator to Lokalise is the free online SDL translator. In general, SDL is a translation service, including localization. For a certain amount you can order a professional translation there, for example, documentation on the chosen language direction. As a bonus, SDL keeps its own online machine translator, which is connected to the Lokalise interface.
Already familiarly check how the SDL will cope with the set of words used to sign the buttons in the application:
At first glance, it is clear that SDL is the “machine-made” translator of all four. It preserves the word order of the source language in the case of phrases. Surprisingly, the long-suffering "Home" SDL translated as "Home", as well as Google Translate. It is difficult to say more about SDL, since the level of translation of even the simplest phrases can be seen in the screenshot.
Now we’ll arrange for the translator to check our Caroline Miber and her “cheap imitation of the bag with the alligator”:
The only thing the SDL online translator coped with is the “train in the afternoon”. However, the translation form itself violates the “who and where” causation, as a result, completely changing the meaning of the sentence - in the three previous cases this did not happen. From the middle of a sentence some absurdity begins. By the example of this sentence, it is clear that the SDL is adapted only to a direct, like an arrow, monosyllabic translation, and the actual constructions and sentences yield something of the level of the legendary “Promtovsky” translation of the middle of zero.
Three of the four online translators connected to Lokalise more or less adequately deal with both monosyllabic constructions and complex sentences of American classics. On a set of indicators ahead, Google Translate is expected, but it does not always work out completely and completely correctly. Microsoft's Bing translator proved to be much better than expected. Apparently, the reason is the use of neural networks. In some aspects, Bing translated even better than Google Translate.
It should be extremely neat to the translator "Yandex". It is clear that the company does not have such resources as Google and Microsoft, but against them the translator of the Russian company looks weak, especially considering that the game was played in his own field - we translated from English to Russian. Why the developers were not able to show the maximum on this one, we are sure, the most popular language pair and to justify marketers' statements about “No. 1 in Russia” is a mystery.
The SDL was not expected to show anything sensible, at least for 2017. This is just a machine translator, not too smart and too straightforward. Perhaps it is suitable for some specific purposes, or simply to double-check a specific word as one of the sources, but no more.
Summing up, I would like to say that in order to get a high-quality translation, which you will not have to return to and redo, you will have to use all possible tools. Of course, the best option would be to attract a living person, but if you are limited in resources, be it time or money, then do not hesitate to use the entire “top three finalists” represented by Google, Yandex and even Bing.
Localization platform for developers of mobile and web projects :
Google Machine Translate / Google Neural Translate
About six months ago, Google announced the connection of another set of languages to the neural network of its Google Translate service, including Russian. This event has become a landmark for the entire Russian-language Internet space: every day thousands of people use Google’s built-in Google translator or go to Google Translate to translate a foreign text into their native language.
Google Translate still retains many features of the classic machine translator based on the rules: when translating a word, the service displays, in addition to the main, most acceptable translation variant of the word, and all its meanings as different parts of speech. In one of our past publications, translations, we have already talked about neural machine translation and how it works.
')
Neural translation, unlike statistical rules based on rigid rules, uses a complex and not always obvious vector system of comparisons in order to choose the most suitable translation based on the context. Of course, it’s not always successful, but more often than not, Google Translate is quite successfully coping with its tasks.
Analysis of the context in the translation field in Google Translate is very easy to check. For the test, we introduce a few words and phrases that can be used in the interface of 90% of applications and see what the translator from the search giant offers:
It seems that the task is extremely simple, but the intellectual side of Google Translate is manifested in one small nuance: the translation of the word “Home”. Based on the context, the translator proposed the option “Home”, which comes closest to the meaning, if we are talking about the software interface in terms of use in the open spaces of the network. At the same time, if we eliminate the context in the form of “app settings” and other words, the translation will remain unchanged - “Home”. All other options in the form of "home" and "dwelling" will be offered in the form of a list under the translation form:
To achieve “machine” from Google translate in the usual sense of the word “Home” translation as “Home” or “Home” is possible only by adding the context - the minimum, in the form of the preposition “at”. With the pretext of Google Translate, it is friendly and instantly translates "at home" as "home."
We will give the translator of the search giant a more difficult task, namely the first sentence of Theodor Dreiser’s novel Sister Kerry. In the original, it looks like this:
Sheet upholstered for a long time, for example, caring out her coat of hair with his sister's street, and his four dollars in money.
In the now classic Soviet translation of Mark Grigorievich Volosov (1895-1941), this sentence looks like this:
When Caroline Mieber got on a train that went to Chicago during the day, all her belongings were in a small chest, a cheap suitcase of fake crocodile skin, a breakfast box and a yellow leather wallet where a train ticket lay, a piece of paper with her sister’s address Buren Street, and four dollars.
Here's how Google Translate handled the translation of this not so simple sentence:
When Carolyn Meeber sat down at noon in Chicago, her overall outfit consisted of a small chest, a cheap imitative alligator bag, a small lunch in a paper box and a yellow leather wallet containing her ticket, a piece of paper With her sister’s address on Van Buren Street and four dollars in the money.
Objectively, the result is excellent for the car. In the preferred translation, the network has lost the word "train", but it is present in one of the options. Also, problems arose with "imitation alligator-skin satchel" and "four dollars in money". The rest of the network performed the translation at a more than decent level for understanding, which requires only stylistic processing.
Yandex translator
The Yandex company likes to position itself as the best in everything within the Russian market, and even bypassing the prohibitions, FAS says it is “search number 1”, for example. Whether this statement is true for other products of the company, it will be possible to check at work the built-in translator “Yandex”, which is located even on a domain similar to its elder brother from Google - translate.yandex.ru .
Much has been written about the translator of Yandex, and by the developers themselves. On their own site, the company's developers immediately admit that the translator of Yandex is a statistical one. To make it clearer and not have to go back to the text about the translators mentioned above: in short, statistical translation involves the enumeration of all possible options with the calculation of the most appropriate. You can also bring the whole slide that the team drew to more clearly show the way the Yandex translator works:
This translator initially loses Google Translate, because it does not have a self-learning neural network. However, the Yandex team had at least one trump card: they could try to make the translator’s work on the “English-Russian” pair as correct as possible, since, obviously, the main target audience of the product is Russian-speaking users.
So, let's feed Yandex to the Translator the same set of words with which we initially tested Google Translate:
As a result, we have two misses: "Options" instead of the expected "Options", "Home" instead of "Home". At the same time, the translator persists and offers the translation of “Home” in the context of “Home” only in conjunction with the word “menu”. In other translation variants, the word “Home”, the variant from “Main”, so conveniently slipped into Google Translate based on the needs of the Internet audience, is completely dropped somewhere to the very bottom. The translator offers only the generic version of "House", "House" and so on.
Now let's see how Yandex machine translation will cope with Dreiser’s work:
When Caroline Meeber boarded a train for Chicago in the evening, her overall outfit consisted of a small trunk, a cheap imitation crocodile skin bag, a small lunch in a paper box, and a yellow snap leather wallet containing her ticket, a scrap of paper from her sister's address in Van Buren Street, and four dollars in money.
In the long run in the form of complex constructions, the problems of older technology become more pronounced. Yes, a few years ago, Google Translate translated the same way, but now the difference, as they say, is “obvious”.
The Yandex translator is experiencing the same problems that people who are learning Russian are faced with: it translates too literally, using less context, and does not understand the cases. In general, cases are the pain of any foreigner. We, as native speakers, who have been practicing the proper use of cases since childhood, do not understand how difficult it really is. We think that there is no sense in analyzing all the errors of the Yandex translator: the nature of the machine statistical translation is very clearly shown in the example above. The only thing with which the domestic child did a great job was the form of recording the name of the street.
Yes, the developers say they use both machine learning and dictionaries. However, it is obvious that for processing complex texts or understanding the context, this is not enough.
Microsoft Translator - Bing Translator
The next on the list of online translators connected to Lokalise is a product of Microsoft, which is used by their search engine Bing. By the way, the Bing translator is more popular in the world than you can immediately think of: it is embedded in all official applications for Windows Phone (for example, in Twitter), used with the default search in Edge for Windows 10, “hooked” to Skype.
Another distinctive feature of Bing is the use of neural networks. First, the neural networks were hooked up to the Bing search to make it more intelligent and expand the page indexing table, which was two times smaller than Google’s. This happened in early 2015. Later, neural networks reached the Bing translator . Thus, it can be argued that the Bing translator is practically in the same "weight category" with Google Translate.
As before, we first check how Bing Translator will work out a simple set of words of interest:
In general, the translator from Microsoft did better than the system from Yandex, but at the same time he allowed himself a lot more liberties than Google Translate. In the case of our stumbling block - the word "Home" - Bing tried to get out and offered the option "Home", which is associated with the button "Home", but certainly not with "Home" and "Dwelling" in the case of "Yandex". On the other hand, the controversial translation of “Tools” as a “means” raises some questions. At the same time, the Yandex machine translator unequivocally translated it as “Tools”. Perhaps Bing is even too creative, so it’s worth checking out with Dreiser:
When Caroline Miber was boarding an afternoon train in Chicago, her overall equipment consisted of a small trunk, a cheap imitation bag with an alligator, a small lunch in a paper box and a yellow leather bag containing her ticket, a piece of paper with her sister’s address on Van Buren Street, and four dollar in money.
In defense of Bing Translator, you can immediately say that so far from all three translators, only he correctly grasped the essence of the phrase “the afternoon train”. The Yandex translator made this afternoon train an evening train, and Google Translate pretended to “pretend to be a hose” and ignored the existence of the train itself. Like his brethren, Bing stumbled over a “cheap imitation bag with an alligator”, and also messed up with “equipment” and “trunk”.
In general, we can say that Bing did a little better than Yandex, but Google Translate translates difficult constructions very far to Google. Probably, there is a lack of sampling and examples for the Bing neural network, since Google translator is obviously much more popular and has great resources.
SDL Free Translator
The latest translator to Lokalise is the free online SDL translator. In general, SDL is a translation service, including localization. For a certain amount you can order a professional translation there, for example, documentation on the chosen language direction. As a bonus, SDL keeps its own online machine translator, which is connected to the Lokalise interface.
Already familiarly check how the SDL will cope with the set of words used to sign the buttons in the application:
At first glance, it is clear that SDL is the “machine-made” translator of all four. It preserves the word order of the source language in the case of phrases. Surprisingly, the long-suffering "Home" SDL translated as "Home", as well as Google Translate. It is difficult to say more about SDL, since the level of translation of even the simplest phrases can be seen in the screenshot.
Now we’ll arrange for the translator to check our Caroline Miber and her “cheap imitation of the bag with the alligator”:
When Caroline Meeber took the afternoon train to Chicago, her costume consists of small groups of SLs, cheap imitation of alligator leather, pounds of plastic small lunch in a field of paper and yellow leather spring retainer bag containing her ticket, a piece of paper from her sister's address in Van Buren Street, and four dollars in money.
The only thing the SDL online translator coped with is the “train in the afternoon”. However, the translation form itself violates the “who and where” causation, as a result, completely changing the meaning of the sentence - in the three previous cases this did not happen. From the middle of a sentence some absurdity begins. By the example of this sentence, it is clear that the SDL is adapted only to a direct, like an arrow, monosyllabic translation, and the actual constructions and sentences yield something of the level of the legendary “Promtovsky” translation of the middle of zero.
Total
Three of the four online translators connected to Lokalise more or less adequately deal with both monosyllabic constructions and complex sentences of American classics. On a set of indicators ahead, Google Translate is expected, but it does not always work out completely and completely correctly. Microsoft's Bing translator proved to be much better than expected. Apparently, the reason is the use of neural networks. In some aspects, Bing translated even better than Google Translate.
It should be extremely neat to the translator "Yandex". It is clear that the company does not have such resources as Google and Microsoft, but against them the translator of the Russian company looks weak, especially considering that the game was played in his own field - we translated from English to Russian. Why the developers were not able to show the maximum on this one, we are sure, the most popular language pair and to justify marketers' statements about “No. 1 in Russia” is a mystery.
The SDL was not expected to show anything sensible, at least for 2017. This is just a machine translator, not too smart and too straightforward. Perhaps it is suitable for some specific purposes, or simply to double-check a specific word as one of the sources, but no more.
Summing up, I would like to say that in order to get a high-quality translation, which you will not have to return to and redo, you will have to use all possible tools. Of course, the best option would be to attract a living person, but if you are limited in resources, be it time or money, then do not hesitate to use the entire “top three finalists” represented by Google, Yandex and even Bing.
Localization platform for developers of mobile and web projects :
Source: https://habr.com/ru/post/336566/
All Articles