When is a translation considered a machine translation?

[tahu157]

Obviously when you're copy and pasting entire sentences and using whatever comes out on the other side, that's a machine translation, even if you "clean it up a bit to make it readable." That's not really what I'm asking about.

What I'm wondering about is translations in which the translator has a decent grasp on Japanese sentence structure, flow, and grammar, but doesn't have a huge amount of vocabulary memorized. So much so that practically every word is run through Jisho for a definition. Is that still considered a machine translation? I'm kinda leaning towards saying yes, that's a machine translation, but I'm curious as to what other people think.

Also, if excessive Jisho use really is still considered a machine translation, how much machine do you have to subtract from the equation to make it a human translation?

[1P1A]

7 minutes ago, tahu157 said:

What I'm wondering about is translations in which the translator has a decent grasp on Japanese sentence structure, flow, and grammar, but doesn't have a huge amount of vocabulary memorized. So much so that practically every word is run through Jisho for a definition. Is that still considered a machine translation?

This is definitely machine translation. Jisho only gives English approximations (cognates don't really exist) meant for beginner/intermediates to give them a foothold on the language. You also can't really substitute experience with machines anyway.

10 minutes ago, tahu157 said:

Also, if excessive Jisho use really is still considered a machine translation, how much machine do you have to subtract from the equation to make it a human translation?

You translate when you're ready, which means you read it and understand it. Ultimately, you shouldn't have to look up much of anything, and those you do, you generally would be able to understand just from the context anyway.

Just my 2 cents.

[ittaku]

30 minutes ago, 1P1A said:

You translate when you're ready, which means you read it and understand it. Ultimately, you shouldn't have to look up much of anything, and those you do, you generally would be able to understand just from the context anyway.

Japanese has a super rich vocabulary. Even the most confident translators would often look up words in translations, sometimes just for confirmation, but not infrequently simply because they've never encountered the word before or it's being used in a context that they've never seen it used before thus confirming what other possible uses of the word there may be. From what I've read, most tertiary educated native speakers know approximately 50,000 words. Literature, depending on complexity, often uses infrequently used words for colour, and even native speakers will encounter words they've never seen before - often they can surmise the meaning simply from context simply because it is their native language, but not always. Guessing at a word's meaning is not accurate enough for translation. By the time one is fluent in a second language, they usually know about 10,000 words. The likelihood of needing a dictionary is still high at that level, yet they clearly have the knowledge to translate. The most famous Japanese dictionaries have almost 500,000 words in them...

[tahu157]

Followup question: in your estimation, what percentage of fan patches are a bunch of Jisho definitions strung together?

[1P1A]

1 minute ago, ittaku said:

Japanese has a super rich vocabulary. Even the most confident translators would often look up words in translations, sometimes just for confirmation, but not infrequently simply because they've never encountered the word before or it's being used in a context that they've never seen it used before thus confirming what other possible uses of the word there may be.

Not disagreeing with this at all. I was more generalizing towards your everyday VN, as opposed to something requiring a high linguistic ability. I never said that you never need to refer to a dictionary, just that when you do you should already have a general idea of what the word/structure itself is symbolizing given the context (again referring to a general VN).

[ittaku]

2 minutes ago, 1P1A said:

Not disagreeing with this at all. I was more generalizing towards your everyday VN, as opposed to something requiring a high linguistic ability. I never said that you never need to refer to a dictionary, just that when you do you should already have a general idea of what the word/structure itself is symbolizing given the context (again referring to a general VN).

Agree with you, I do.

 

[phantomJS]

1 hour ago, tahu157 said:

Followup question: in your estimation, what percentage of fan patches are a bunch of Jisho definitions strung together?

List of fan-translated VNs that I had read so far: Dracu Riot, Nobel Works, Majikoi series, Hatsukoi 1/1, Rewrite, Tsujidou-san, Eiyuu Senki, Fate Stay/Night, Fate Hollow, Kono Oozora, Koi to Senkyo to Chocolate.

The TL in all of them are generally excellent in quality and aren't a bunch of dictionary definition mashed together. Although my list is relatively small, I believe it should be pretty representative of the entire fan-TL scene in general. Personal gut feeling is those who are weak in Jap will not come and work on a TL for free as TL-ing requires passion, commitment, and confidence to do so. This in turn will keep Fan-TLs high in quality.

It's the 'professional' TL scene which you should ask and worry about. Seriously, the crap job Moenovel did with Kono Oozora and Cross Channel is face palm worthy

Edited May 7, 2018 by phantomJS
Forgot 1 VN; Koi Choco

[Darklord Rooke]

55 minutes ago, phantomJS said:

Fate Stay/Night

The TL in all of them are generally excellent in quality

[Clephas]

1 hour ago, phantomJS said:

List of fan-translated VNs that I had read so far: Dracu Riot, Nobel Works, Majikoi series, Hatsukoi 1/1, Rewrite, Tsujidou-san, Eiyuu Senki, Fate Stay/Night, Fate Hollow, Kono Oozora, Koi to Senkyo to Chocolate.

The TL in all of them are generally excellent in quality and aren't a bunch of dictionary definition mashed together. Although my list is relatively small, I believe it should be pretty representative of the entire fan-TL scene in general. Personal gut feeling is those who are weak in Jap will not come and work on a TL for free as TL-ing requires passion, commitment, and confidence to do so. This in turn will keep Fan-TLs high in quality.

It's the 'professional' TL scene which you should ask and worry about. Seriously, the crap job Moenovel did with Kono Oozora and Cross Channel is face palm worthy

MirrorMoon's fantranslations are pretty grammatically-challenged, from what I recall... and I can say from working on fixing it that Dracu-riot (the version you must have played) had a lot of problems.   I'd say about one third of all fantranslations have a translator who is constantly referring to his dictionary and has trouble grasping context clues... or is just grammatically challenged (even if his Japanese understanding is just fine).  All of the Eushully games that have patches are basically 'modified machine translations'.

The professional scene has gotten better in some ways, worse in others.  SP and Mangagamer are producing a consistent level of quality these days... but I still recall the original localizations of Edelweiss, Da Capo, and Suika (think ATLAS machine translations with no attempt to fix the grammar).  These days, we are more likely to see that kind of crap out of new companies than the older ones (though Moenovel seems to be incapable of doing anything right) as the Libra and Chuusingura translations show.

[phantomJS]

@Darklord Rooke @Clephas

I was comparing them to 'professional' releases.

Even the TL of FS/N looks great when compared to the likes of Libra of the Vampire Princess, Edelweiss, and most things from Moenovel. I think the TL quality of MG releases is pretty average; I got Princess Evangile W Happiness on the day it came out and the TL was disappointing considering they took around 2 years to TL it. Magical Marriage Lunatics too was mediocre. When looking at it in this context, I think the stuff the Fan-TL scene puts out is great.

Anyway, just my thoughts

Edited May 7, 2018 by phantomJS

[Dergonu]

5 hours ago, tahu157 said:

Followup question: in your estimation, what percentage of fan patches are a bunch of Jisho definitions strung together?

One important thing to keep in mind is, there is a big difference between understanding everything in a sentence on a grammatical level, but having to look up one word, (like a noun,) to fill in the blanks, and having to look up every single part of the sentence. If all you are lacking to understand a sentence fully is one word, then you are still translating it like a professional would have after you learn said word. All you have to do is look up one word, after all, something that you might want to do even if you know a word, just to make sure you got the usage down right.

I personally use a Japanese dictionary on the side of Jisho, in order to make sure I have the Japanese contextual usage down right, and often end up looking up words I already do know, just to make sure. The best way to understand how a word/ phrase is used is to read the Japanese definition and Japanese example sentences.

So, "a string of jisho definitions" would, in my head, just mean someone who did not understand the sentence on a grammatical level, and ended up looking up everything, which in turn means it's a machine translation. But if you're asking how many fan TLers use jisho to look up words frequently to either learn them, or make sure they got the word's definition down, then I'd say all of them. (Even professional TLers do that, after all.)

[Infernoplex]

11 hours ago, Clephas said:

All of the Eushully games that have patches are basically 'modified machine translations'.

Even the Aroduc's TL of Kamidori?

[kokoro]

31 minutes ago, Infernoplex said:

Even the Aroduc's TL of Kamidori?

He too.

[tahu157]

9 hours ago, Dergonu said:

So, "a string of jisho definitions" would, in my head, just mean someone who did not understand the sentence on a grammatical level, and ended up looking up everything, which in turn means it's a machine translation. But if you're asking how many fan TLers use jisho to look up words frequently to either learn them, or make sure they got the word's definition down, then I'd say all of them. (Even professional TLers do that, after all.)

I was thinking more the former than the latter. I think it's /technically/ possible to know how to put all the pieces of a translation together accurately even if you have to use software to acquire all the pieces. That said, I also think that if you're at a level where you have to get assistance for all the pieces you probably also don't have as good a handle on Japanese grammar as you might think.

That's basically where I am right now, finding easy things to read with the assistance of Jisho for basically every word I encounter. It's a pretty frustrating process but I'd also say it's the most effective learning method I've tried thus far.

[Clephas]

2 hours ago, Infernoplex said:

Even the Aroduc's TL of Kamidori?

Everyone at the time knew that he had basically modified a machine translation for Kamidori.  He just did a slightly better job than most people would have.

Edit: To be clear, machine translations were not as much of a stigma back then as they are now.  The main reason it became such a stigma was because of the early Mangagamer efforts (which were abominable).

[Infernoplex]

17 minutes ago, Clephas said:

Everyone at the time knew that he had basically modified a machine translation for Kamidori.  He just did a slightly better job than most people would have.

Edit: To be clear, machine translations were not as much of a stigma back then as they are now.  The main reason it became such a stigma was because of the early Mangagamer efforts (which were abominable).

From what I have seen of it, it didn't look bad to me (granted, I didn't fully play it, but I did look at early parts of it). If that was based on an MTL, then it was a very good edited MTL because I would never tell that was based on an MTL.

[Clephas]

49 minutes ago, Infernoplex said:

From what I have seen of it, it didn't look bad to me (granted, I didn't fully play it, but I did look at early parts of it). If that was based on an MTL, then it was a very good edited MTL because I would never tell that was based on an MTL.

He did a very, very good job at editing it (very thorough).  The reason machine translations became such a huge stigma later on was because of later abortive efforts by others, including the aforementioned incidents with Edelweiss, where the machine translations were basically presented without any grammatical fixes at all.  Of course, what he did was a full TLC/edit, but the fact is he based it off a machine translation. 

I played it, and while - from a purely translation point of view - it has issues, as reading material, it is mostly just fine.  He ran extensive fixes for internal consistency, and he made a serious effort to make it readable in general.  A lot of the reason why it was so successful was because Kamidori, in general, is not one of the harder Eushully games linguistically (I wouldn't want to see him make the same effort with IMZ, for instance).  There is little attempt at the kind of large-scale world-building you see in IMZ or the later games by Eushully, and the story itself is relatively straightforward and is not a 'heroic epic' like the IM games, so the narration is more simplistic.

Understand, the reason most translators bash attempts at machine translation (other than elitism) is because the people who are using them don't have the Japanese or the English skills necessary to know when the output they are getting is wrong and/or needs modification.  That wasn't the case with the Kamidori fantranslation...  and at the time, nobody really cared it was based off of MTL.  After all, it was a readable VN at a time when the localization industry was basically JAST and the nascent MG.

[Infernoplex]

13 minutes ago, Clephas said:

He did a very, very good job at editing it (very thorough).  The reason machine translations became such a huge stigma later on was because of later abortive efforts by others, including the aforementioned incidents with Edelweiss, where the machine translations were basically presented without any grammatical fixes at all.  Of course, what he did was a full TLC/edit, but the fact is he based it off a machine translation. 

I played it, and while - from a purely translation point of view - it has issues, as reading material, it is mostly just fine.  He ran extensive fixes for internal consistency, and he made a serious effort to make it readable in general.  A lot of the reason why it was so successful was because Kamidori, in general, is not one of the harder Eushully games linguistically (I wouldn't want to see him make the same effort with IMZ, for instance).  There is little attempt at the kind of large-scale world-building you see in IMZ or the later games by Eushully, and the story itself is relatively straightforward and is not a 'heroic epic' like the IM games, so the narration is more simplistic.

Understand, the reason most translators bash attempts at machine translation (other than elitism) is because the people who are using them don't have the Japanese or the English skills necessary to know when the output they are getting is wrong and/or needs modification.  That wasn't the case with the Kamidori fantranslation...  and at the time, nobody really cared it was based off of MTL.  After all, it was a readable VN at a time when the localization industry was basically JAST and the nascent MG.

I see. Thanks for the explanation. I didn't originally know the story behind it, but I do know that other Eushully stuff got only interface patches and pure MTL for the story (which doesn't read anything like it should). Compared to those, it's a breath of fresh air and the only Eushully VN fantranslation worth playing. As for early MG stuff, I only read a couple of stuff they localized like the "Suck my Dick or Die", "Suika", "My Sex Slave is a Classmate" and... "Edelweiss", although not the first version but the second one. The worst among these was the classmate one. Now that's completely atrocious, both content and translation-wise. I heard about Suika's issues later on, but when I was reading it... it didn't look that bad at all as they make it out to be. Definitiely better than the poor classmate or that other Liquid nukige. Edelweiss read fine, at least the second version I grabbed. It wasn't 100% perfect, but it was good enough for me and I liked it.

Edited May 7, 2018 by Infernoplex

[tymmur]

16 hours ago, tahu157 said:

What I'm wondering about is translations in which the translator has a decent grasp on Japanese sentence structure, flow, and grammar, but doesn't have a huge amount of vocabulary memorized. So much so that practically every word is run through Jisho for a definition. Is that still considered a machine translation? I'm kinda leaning towards saying yes, that's a machine translation, but I'm curious as to what other people think.

I would say no. Say a word can be translated into 4 different words. The translator will read the definition of each case and pick the best suited. MT will pick a semi random. That alone makes it worth telling human translators apart from MT. For a human made translation to be considered as a machine translation, the translator should do something stupid like always picking the first choice for each word.

 

As for dictionaries, I don't really have a problem with people using them. In fact it would be a far worse problem if people don't use them at all. Translation is not about just knowing as many words as possible. In fact that's the least of the problems. The importance of grammar can't be overestimated because screwing up grammar will likely lose a lot of the meaning even if you get the words right. Another issue is how each language handles how to say different concepts. We had one in an introduction thread recently, which was "native language" vs "language of your mother". Other examples are homeland vs fatherland vs motherland and "they shook their heads" vs "they shook their head" (first is English with logical plural as it counts the total number of heads. Some languages count the number on each person, meaning plural will make multi headed people). Those are examples of translation issues between Germanic languages, despite being closely related. Now take two completely unrelated languages like English and Japanese and differences like that will be in a significantly higher number. By the time a person master issues like that, the vocabulary will most likely be good enough for that to not be an issue.

 

Another note regarding dictionaries is that they have to be used often during translation work. The accuracy required for a good translation mean if in doubt, look it up. If it looks a little bit out of the ordinary, it could be a double meaning, a play on words, kanji or similar. Even native speakers can have a hard time picking up on all the details, which should hopefully be included in a translation.

3 hours ago, tahu157 said:

That's basically where I am right now, finding easy things to read with the assistance of Jisho for basically every word I encounter. It's a pretty frustrating process but I'd also say it's the most effective learning method I've tried thus far.

I would say it's likely the best approach to build up a vocabulary because you only encounter words, which are actually used and if you encounter the same word again and again, you will have a higher chance of remembering it, meaning your learning is optimized for learning commonly used words. Some people write words on cards (Japanese on the front and translation on the back) and then try to remember each morning before flipping. This is sometimes done with random words from the dictionary and even if they are successful, it will be completely useless because they often end up using words they will never encounter. It might work for learning hiragana or something, but not really to get a useful vocabulary.

[sanahtlig]

The definition of "machine translation" being used and accepted in this thread seems odd to me.  A machine translation, by definition, lacks knowledge of context.  If a human accurately fills in that context, it's no longer a machine translation.  The problem is when an editor, with no knowledge of Japanese, comes along and inaccurately interpolates ambiguous / wrong output.  Coincidentally, a translator fluent in Japanese can make context mistakes too, especially if they're translating from a script without ever having played the game (and missing the cues in the visuals and audio that provide added context).

[Yuuko]

https://en.wikipedia.org/wiki/Computer-assisted_translation

[tahu157]

1 hour ago, sanahtlig said:

The definition of "machine translation" being used and accepted in this thread seems odd to me.  A machine translation, by definition, lacks knowledge of context.  If a human accurately fills in that context, it's no longer a machine translation.  The problem is when an editor, with no knowledge of Japanese, comes along and inaccurately interpolates ambiguous / wrong output.  Coincidentally, a translator fluent in Japanese can make context mistakes too, especially if they're translating from a script without ever having played the game (and missing the cues in the visuals and audio that provide added context).

While I do agree that human-interpreted context added to individual machine-generated definitions doesn't fit the traditional machine translation definition, I would argue that that method is still more machine than man, so to speak. Maybe the better question to ask is whether or not someone that has to refer to a machine-dictionary excessively is actually capable of accurately interpreting the context of a Japanese sentence. Would that end up being better, or as bad as a pure/edited machine translation?

1 hour ago, Kiriririri said:

https://en.wikipedia.org/wiki/Computer-assisted_translation

How insightful.

 

[1P1A]

48 minutes ago, tahu157 said:

Maybe the better question to ask is whether or not someone that has to refer to a machine-dictionary excessively is actually capable of accurately interpreting the context of a Japanese sentence. Would that end up being better, or as bad as a pure/edited machine translation?

You can use the "translation" of Venus Blood Chimera (not to be confused with the currently being localized Venus Blood title) as an example. Apparently the fellow just put together words from a parser without even having the basic understanding of particles. From some of the examples posted about 4chan/vndb, it looked pretty fan-fictiony.

From excessively referring to a dictionary they are focusing more on the words, and not what the words themselves are alluding to. Sure, you can grammatically put stuff together and try to make sense of it (even professionals have to do this sometimes), but you'll end up missing a lot if that's the main focus. Translation takes a long time, and if you have to dissect every sentence that just compounds.

If someone has to reason why they should be allowed to translate, it's probably better to focus on the language instead. Translating and learning Japanese only overlap on a small level, and no large amount of time should be used if that's a main goal.

[tymmur]

1 hour ago, tahu157 said:

Maybe the better question to ask is whether or not someone that has to refer to a machine-dictionary excessively is actually capable of accurately interpreting the context of a Japanese sentence. Would that end up being better, or as bad as a pure/edited machine translation?

I once did a test to figure out something like that. Way back before I read any Japanese grammar, I started translating 20 lines using TA+jparser. I had a printed paper with all the particles with each explained and then the dictionary lookup from jparser and that's it. I had set furigana to romaji as well (hiragana skills should not influence the test result). The goal was not to see if it would be possible to translate VNs like that, but rather to figure out what the threshold is to beat MT in quality when you read yourself.

To answer the question: human beats machine in sentences with simple grammar while machine has the edge if the sentence use more advanced grammar. Now if the human had prepared and studied something like Tae Kim's grammar guide (free and in English) prior to starting, then MT would lose big time.

However it's important to be aware that results highly depends on the person doing the translation.

EDIT: disclaimer: I did not work on very complex sentences. It would be pointless to look at something where we would expect both human and machine would fail. As such, simple grammar really do mean really simple grammar.

4 minutes ago, 1P1A said:

Translating and learning Japanese only overlap on a small level, and no large amount of time should be used if that's a main goal.

Translating forces the translator to focus on all the details in the text, which could be skipped as a reader. This mean translating could be considered some sort of language training and it will also increase the vocabulary. If used correctly, it could be a tool to learn Japanese, but it's not a replacement for studying. It's more like if you use what you read in the schoolbook, then you have a better chance of remembering it. Also getting some familiarity with Japanese mean if you read about a new concept, you can go "ohh, I saw something like this already. It was that odd thing the heroine said that I didn't fully get. Now I get it". Something like that can really benefit in remembering what you learn from a book. Translation work is however not a replacement for studying.

It's also important to note that people are different. There is not one single best approach to learn Japanese because the best approach depends on the person. What looks well for one person might not work as well for somebody else.

Edited May 8, 2018 by tymmur

[1P1A]

30 minutes ago, tymmur said:

Translating forces the translator to focus on all the details in the text, which could be skipped as a reader. 

I agree with this and is what I had in my mind when I wrote my reply.

34 minutes ago, tymmur said:

If used correctly, it could be a tool to learn Japanese, but it's not a replacement for studying. Translation work is however not a replacement for studying.

I was hoping my post would have conveyed this, but it's good to have it stated concretely.

35 minutes ago, tymmur said:

It's also important to note that people are different. There is not one single best approach to learn Japanese because the best approach depends on the person. What looks well for one person might not work as well for somebody else.

Very good point. The best form of studying is the one that keeps you motivated to keep going, but doing something that's arduous, and shows little to no results, probably isn't going to do that. But I can totally see the allure of treating the language like a puzzle to be solved.