Jump to content
Slate Blackcurrant Watermelon Strawberry Orange Banana Apple Emerald Chocolate Marble
Slate Blackcurrant Watermelon Strawberry Orange Banana Apple Emerald Chocolate Marble
NowItsAngeTime

A comparison of using Machine Translations (VNR + Google Translate) vs Application of Grammar/Vocab Knowledge for Reading Visual Novels in Japanese

Recommended Posts

Intro

This is just a random idea I had.

So every so once in a while I see people who are interested in reading Japanese-only visual novels and consider using something like Visual Novel Reader or some kind of other way to machine translate the VN for them. I then see people who respond that machine TLs are bad and that they should just learn Japanese.

As someone who's started learning Japanese last year and have been making ok progress, I wanted to show the differences of how someone learning Japanese could read a sentence based on comprehension vs. how a machine translator like Visual Novel Reader (using LEC and ATLAS dictionaries) and Google Translate would translate the same sentence.

I'm not gonna sit here and be one of those JOP shills who say JUST LEARN JAPANESE. I'm just gonna provide the analysis and comparison and let the people decide if the differences don't matter to them or if they want to get inspired to learn Japanese if they really want to. Remember, learning a second language isn't easy and you gotta put in the work which takes time and practice and sacrifice.

----------------------------------------------------------

Details about the Demonstration

The Visual Novel I'm going to use for this demonstration is my first visual novel I'm currently reading in Japanese: Kanojo to Ore to Koibito to. This is a slice of life moege so it should be easier to digest than a visual novel with a lot of plot.

I'm not gonna translate a whole scene because that would take multiple posts. Instead, I'm taking 10 lines of a simple conversation and posting them in the order they appear.

For the purposes of my interpretations, I'm going to be using a combination of Jisho and Weblio to look up vocabulary and Taekim's Grammar Guide and other sites to confirm proper grammar usage.

I'm not going to give an explanation for every single personal translation, but if you're a seasoned Japanese reader and want to correct me on things that look wrong, feel free as I'm still learning.

I'll also be pointing out any positives or negatives of the machine TLs. I'm not just going to shit all over machine TLs because I want to see what positives they may offer.

Context for this scene (no spoilers, this is the third scene in the whole VN)

(Note: Context is VERY important for translating/reading properly in Japanese. Which you will soon see.)

The day before, Haruto saved Ayano when he jumped out in front of an errant vehicle that was heading for her, but he got hurt in the process.

He got put in the hospital because of it and Ayano visited him and they basically hit it off. Note they had never met before this occasion. She promised to keep visiting him every day until he got better.

The scene below is the day after Haruto got put in the hospital. Ayano just knocked on the door to his room.

Another thing to note is that Haruto has a more casual way of talking, while Ayano is much more polite. Polite vs casual speech is another thing to take note of in the language.

Full Names:

  • 松上 遥人 Matsugami Haruto
  • 美萩野 綾乃 Mihagino Ayano

----------------------------------------------------------

Line 1

Haruto -「はい、どうぞー」

Romaji: Hai, douzoー

1 to 1 translation: 「Yes, go aheadー」

Personal Translation: 「Come inー」

Visual Novel Reader, LEC: 「Yes. Certainly」

Visual Novel Reader, ATLAS: 「Yes,Please」

Google Translate: "Yes, please."

Analysis:

So どうぞ is one of many words that have potential multiple meanings: https://jisho.org/search/douzo.

It's a simple word that depending on context either means "feel free to ___" or "please" or "Here you are".

The machine translated sentences clearly used the meaning that goes against the context of the situation. Just because the meanings they chose are probably more common, doesn't mean it's the go to meaning to use for all situations.

----------------------------------------------------------

Line 2

Ayano - 「失礼いたします」

Romaji:「Shitsurei itashi masu」

1 to 1 translation: 「Excuse me.」

Personal translation: 「Excuse me.」

Visual Novel Reader, LEC: 「I leave」

Visual Novel Reader, ATLAS: 「I will be must be going.」

Google Translate: "excuse me"

Analysis:

失礼いたします (or 失礼します) is a common phrase for entering or leaving rooms or conversations/people. LEC and ATLAS seemed to assume it was for someone leaving. Once again, it may be the more common usage of the word but does not mean it's the only usage of it.

Also, the English grammar usage in some of the machine translations are pretty funny to look at.

----------------------------------------------------------

Line 3

Narration - 入ってきたのは、花束を持った美萩野さんだった。

Romaji: Haittekita no wa, hanataba o motta Mihagino-san datta.

1 to 1 translation: Coming in, bouquet held Mihagino-san was.

Personal translation: It was Mihagino-san coming in holding a bouquet.

Visual Novel Reader, LEC: It was Mihagino with the bouquet to have entered

Visual Novel Reader, ATLAS: It was Mr. Mihagino that had entered who had had the bouquet.

Google Translate: It was Mr. Mikagino with a bouquet that came in.

Analysis:

DID YOU JUST ASSUME HER GENDER? As you can see, a few cases of machine translation will see the "san" honorific and automatically assume it's a guy. In addition, apparently google translate isn't always good for showing names.

Spotty English grammar and gender aside, the machine TLs actually covered the main point of the sentence pretty well.

----------------------------------------------------------

Line 4

Haruto - 「あ……っと、美萩野さん」

Romaji: 「A...tto, Mihagino-san」

1 to 1 translation: 「Ah... Oh, Mihagino-san」

Personal translation: 「Ah... Oh, it's Mihagino-san.」

Visual Novel Reader, LEC: 「...Mihagino」

Visual Novel Reader, ATLAS: 「Mr. Mihagino」

Google Translate: "Ah ... ... well, Mr. Mikagino"

Analysis:

More gender assuming. Also it looks like machine translators aren't good with single characters that seem out of place. And Google still doesn't like names.

----------------------------------------------------------

Line 5

Ayano - 「お見舞いにうかがったのですが、お邪魔ではありませんか?」

Romaji: 「omimai ni ukagatta no desuga, ojyama de wa arimasenka?」

1 to 1 translation:「Calling on someone who is ill to visit expected that, interrupt to be not?」

Personal translation: 「I'm visiting you since you were sick. I hope I'm not disturbing you?」

Visual Novel Reader, LEC: 「I asked a visit, but to not disturb it?」

Visual Novel Reader, ATLAS: 「It will be obstructive through looked to the visit.」

Google Translate: ""I was caught asking you, is not you in the way?""

Analysis:

This is a case where context in the Japanese language is very important. Since pronouns aren't used for most sentences, the listener has to infer what or who is being referred to in the dialogue.

It looks like machine translators either guess which pronouns to use or don't even use personal pronouns at all.

----------------------------------------------------------

Line 6

Narration - 驚きが顔に出てたのか、美萩野さんは申し訳なさそうに聞いてくる。

Romaji: Odoruki ga kao ni deteta no ka, Mihagino-san wa moushiwakenasazou ni kitekuru.

1 to 1 translation: Surprise face made?, Mihagino-san apologetically asks.

Personal translation: I must have made a surprised face as Mihigano asked that apologetically.

Visual Novel Reader, LEC: Mihagino hears whether surprise was reflected on a face apologetically.

Visual Novel Reader, ATLAS: Did the surprise appear to the face? Mr. Mihagino hears it seeming to apologize.

Google Translate: Did Mr. Mikagino ask a little sorry for surprise appearing on the face?

Analysis:

In Japanese grammar it's possible to ask a question in a larger sentence that may not end as a question.

It's a bit difficult for me to translate that personally. If the machine translators actually translated the other words in the sentence better, I think they had the right idea with translating embedded questions in the middle of a longer sentence.

----------------------------------------------------------

Line 7

Haruto - 「いや、そんなことない。全然ないからっ」

Romaji: 「Iya, Sonnakotonai. Zenzennai kara」

1 to 1 translation: 「No, that's not true. Not at all not so」

Personal Translation: 「Nah, that's not true. You're not disturbing me at all.」

Visual Novel Reader, LEC: 「No, there is not such a thing. Because there is not it at all.」

Visual Novel Reader, ATLAS: 「Unpleasantness. Such a thing. Because it is not at all.」

Google Translate: "No, that's not true. It is not at all"

Analysis:

"そんなこと" literally means "such a thing" which is what some of the machine translations go for. Generally in actual sentences you mostly won't be saying "such a thing" but rather work around the current topic of the sentence.

As you can see, the negative version can be used to just say something isn't true, so it's another case of a phrase having multiple meanings or interpretations.

----------------------------------------------------------

Line 8

Ayano - 「ありがとうございます。それでは、失礼いたします」

Romaji - 「Arigatou gozaimasu. Soredewa, Shitsurei itashi masu.」

1 to 1 translation: 「Thank you very much. Well then, excuse me.」

Personal Translation: 「Thank you very much. Well then, excuse me.」

Visual Novel Reader, LEC: 「Thank you. Then I leave」

Visual Novel Reader, ATLAS: 「Thank you. Thank, I will must be going.」

Google Translate: "Thank you very much, then I will excuse you."

Analysis:

Another case of 失礼いたします meaning being assumed. At least the pronouns were mostly correct.

----------------------------------------------------------

Line 9

Narration - 持ってきてくれた花を花びんに生けると、前と同じようにベッドの脇の椅子に座った。

Romaji - Mottekitekureta hana o habin ni ikeru to, mae to onajiyou ni beddo no waki no isu ni suwatta.

1 to 1 translation: Brought flowers vase in arrange when, same as before bed's side's chair sat.

Personal Translation: After she arranged the flowers she brought in a vase, she sat on the chair besides the bed like last time.

Visual Novel Reader, LEC: I sat down on a chair beside the bed in the same way as the front when I arranged the flower which I brought in a vase.

Visual Novel Reader, ATLAS: When the brought flower was arranged in the vase, it sat as much ahead on a sideward chairs of the bed.

Google Translate: When I lived on a flower with the flowers that brought me, I sat on the chair next to the bed just as before.

Analysis:

Once again, the general idea is conveyed in the machine TLs but the issue is slight grammar issues as well as assuming pronouns which change the direction of the sentence.

----------------------------------------------------------

Line 10

Haruto - 「ありがとう。来てくれて、嬉しいよ」

Romaji: Arigatou. Kitekurete, ureshii yo

1 to 1 translation: Thank you. Coming, I'm glad.

Personal Translation: Thank you. I'm glad you came.

Visual Novel Reader, LEC: Thank you. I am glad to come

Visual Novel Reader, ATLAS: Thank you. I am glad to come.

Google Translate: "Thank you, I am glad you came"

Analysis:

Google Translate basically got it perfectly in my eyes. LEC and ATLAS unfortunately used pronouns in the incorrect way again which turned the sentence into something more lewd sounding.

----------------------------------------------------------

Conclusion

It seems like the machine translations get the right idea of how to translate the sentence in terms of easier vocabulary and grammar.

However, it seems like it struggles with some major things

1) Words/Phrases with multiple meanings. It takes what's the most common usage of the meaning of the word but since it can't read context, there's a chance what it translates a word/phrase to is incorrect for the situation.

2) Prounouns. As I said, unlike the English language where we basically have to put pronouns everywhere, you don't actually do that a lot of the time in Japanese language. You have to infer it based on the situation. Machine translations can't possibly know the context so it will take a guess.

3) Assuming genders with honorifics.

4) Apparently some names don't get printed out well sometimes. My guess it would be worse with characters that have names that would actually translate to English well like "小鳥" would become "small bird" instead of "Kotori".

5) Arranging sentences with clean English grammar. Usually the translation has grammar that could work, but is just not the most pleasant/common to read.

Share this post


Link to post
Share on other sites

It seems that MTLs do well if sentence is "full textbook" one - with all pronouns, subjects, objects etc present. Once things get omitted, shortened and need to be guessed from context - they fail, or rather they get random.

Edited by adamstan

Share this post


Link to post
Share on other sites

Couple remarks to experiment conditions. Atlas and LEC are non-neuro machine translation services, so it's unfair to compare them with Google translate. I show comparison with other neuro machine translation services like Bing and Yandex. Another point is that substitutions are must have for difficult proper names. Once you subsitite 小鳥 for Kotori, you'll never see small bird in text. The same goes for Mihagino. There are also some tools to improve machine translation with honorifics like TAHelper and replacement scripts, but I did not explore them personally as I'm perfectly fine to see Mr. before girls names as I know that -san is meant.

----------------------------------------------------------
Line 1
Haruto -「はい、どうぞー」
Personal translation: 「Come inー」

Bing: Yes, please.
Yandex: Yes, please.」
Google Translate: "Yes, please."
----------------------------------------------------------
Line 2
Ayano - 「失礼いたします」
Personal translation: 「Excuse me.」

Bing: I'm sorry.
Yandex: Excuse me.」
Google Translate: "excuse me"
----------------------------------------------------------
Line 3
Narration - 入ってきたのは、花束を持った美萩野さんだった。
Personal translation: It was Mihagino-san coming in holding a bouquet.

Bing: It was Mr. Mihagino who came in with the bouquet.
Yandex: Mihagino had a bouquet of flowers.
Google Translate: Mihagino who had a bunch of flowers came in.
----------------------------------------------------------
Line 4
Haruto - 「あ……っと、美萩野さん」
Personal translation: 「Ah... Oh, it's Mihagino-san.」

Bing: "Oh... Mr. Mihagino.
Yandex: Ah....... Mihagino, Mihagino, Mihagino, Mihagino」
Google Translate: "Ah ...... Uh, Mihagino-san"
----------------------------------------------------------
Line 5
Ayano - 「お見舞いにうかがったのですが、お邪魔ではありませんか?」
Personal translation: 「I'm visiting you since you were sick. I hope I'm not disturbing you?」

Bing: "I was asked to visit, but do not disturb you? 」
Yandex: "I asked for a visit, but is not it a bother you?"」
Google Translate: "I was caught asking you, is not you in the way?"
----------------------------------------------------------
Line 6
Narration - 驚きが顔に出てたのか、美萩野さんは申し訳なさそうに聞いてくる。
Personal translation: I must have made a surprised face as Mihigano asked that apologetically.

Bing: The surprise came out on the face, and Mihagino-San hears it apologetic.
Yandex: Mihagino asked me if there was a surprise in his face.
Google Translate: Mihagino asked to be sorry for a surprise appearing on the face.
----------------------------------------------------------
Line 7
Haruto - 「いや、そんなことない。全然ないからっ」
Personal Translation: 「Nah, that's not true. You're not disturbing me at all.」

Bing: No, that's not true. Not at all. "
Yandex: "No, it's not.Not at all.」
Google Translate: "No, that's not true.
----------------------------------------------------------
Line 8
Ayano - 「ありがとうございます。それでは、失礼いたします」
Personal Translation: 「Thank you very much. Well then, excuse me.」

Bing: "Thank you very much. I'll be sorry.
Yandex: "Thank you.So, excuse me.」
Google Translate: "Thank you very much, then I will excuse you."
----------------------------------------------------------
Line 9
Narration - 持ってきてくれた花を花びんに生けると、前と同じようにベッドの脇の椅子に座った。
Personal Translation: After she arranged the flowers she brought in a vase, she sat on the chair besides the bed like last time.

Bing: I sat in a chair on the side of the bed as before, when I put the flowers that I brought in the vase.
Yandex: When I put the flower in the flower vase, I sat on the chair by the bed as before.
Google Translate: When I lived on a flower with the flowers that brought me, I sat on the chair next to the bed just as before.
----------------------------------------------------------
Line 10
Haruto - 「ありがとう。来てくれて、嬉しいよ」
Personal Translation: Thank you. I'm glad you came.

Bing: Thank I'm so glad you're here.
Yandex: "Thank you.I'm glad you're here.」
Google Translate: "Thank you, I am glad you came"
----------------------------------------------------------

Conclusion:

Gender, multiple meanings in context and honorifics remain weak points. There's no arguing that manual translation is better. The actual question is what variants you have to read untranslated works:

1.  JUST LEARN JAPANESE ALREADY! The best variant, but requires either 2+ years of dedicated study with some study group or 5+ years of dedicated study alone, or eternity if you slack.

2. Manual translation. Takes at least x10 times more to read any visual novel, making forget previous events with time and lose enthusiasm with such horrible pace. In theory should help learning Japanese. In practice makes you drowsy each 10 minutes resulting in never finishing anything without proper dedication.

3. Machine neuro translation. Allows you to read whatever you want at a normal pace right away without big hit to the meaning provided you can put up with some broken grammar.

There is no just one right answer. All people are different. First and second variants are suitable only for some 10% of people, others would get demotivated before finishing anything. I advocate for machine neuro translation only because it's available just as you set up your mind to try untranslated visual novels. You get to appreciate the media and then need to decide for yourself if you want to try more works this way first or start learning the language at this point. I also have selfish motives here as I'd like to see untranslated visual novels discussion bloom after all those dark decades for untranslated vns. For those who are 100% sure that reading machine neuro translations automatically makes you understand nothing, I only urge to try machine neuro translation and then read proper English translation. That will make you see for yourself whether there are myriads of hidden meanings lost in initial playthrough or it's the very same story.

Edited by kivandopulus

Share this post


Link to post
Share on other sites

Machine translators can't deal with scene context, implied subjects, and the extended sentences common to Japanese grammar.  In addition, they often misinterpret verb forms and completely mess up the sentence to the point where it is unrecognizable.  This is the reason why only the simplest of sentences that have a clearly delineated subject and use the simplest verb forms actually translate properly (and often not even then).  Before even considering playing untranslated, you should at least know the grammar so you can tell when something is off...  you can fill out your vocab along the way using a parser (puts romaji or kana above kanji). 

If you don't want to learn Japanese, you are out of luck.  Machine translations will be inadequate until we can hook a person with a complete vocabulary and grammar knowledge of both languages up to a machine and download an entire copy of his brain directly into the digital realm (in other words, we won't see it in our lifetimes).  The language just isn't suited for machine translation, especially since there are no languages directly related to it (romance and germanic languages are relatively easy to machine translate to English due to a large swathe of shared concepts and essential structural similarities).  It doesn't even translate well into mandarin or modern Chinese, despite the languages partially sharing a writing system.

Share this post


Link to post
Share on other sites
56 minutes ago, kivandopulus said:

1.  JUST LEARN JAPANESE ALREADY! The best variant, but requires either 2+ years of dedicated study with some study group or 5+ years of dedicated study alone, or eternity if you slack.

2. Manual translation. Takes at least x10 times more to read any visual novel, making forget previous events with time and lose enthusiasm with such horrible pace. In theory should help learning Japanese. In practice makes you drowsy each 10 minutes resulting in never finishing anything without proper dedication.

3. Machine neuro translation. Allows you to read whatever you want at a normal pace right away without big hit to the meaning provided you can put up with some broken grammar.

There is no just one right answer. All people are different. First and second variants are suitable only for some 10% of people, others would get demotivated before finishing anything. I advocate for machine neuro translation only because it's available just as you set up your mind to try untranslated visual novels. You get to appreciate the media and then need to decide for yourself if you want to try more works this way first or start learning the language at this point. I also have selfish motives here as I'd like to see untranslated visual novels discussion bloom after all those dark decades for untranslated vns. For those who are 100% sure that reading machine neuro translations automatically makes you understand nothing, I only urge to try machine neuro translation and then read proper English translation. That will make you see for yourself whether there are myriads of hidden meanings lost in initial playthrough or it's the very same story.

Yes there is just one right answer, and that answer is NOT machine anything translation if you don't know any Japanese.

Share this post


Link to post
Share on other sites

Shouldn't an AI be able to learn proper grammar, use of context and whatnot? It might not be on the table right now for everyday consumers, but I feel like it's something that could really revolutionize translation processes in the near future.

Edited by Seraphim88

Share this post


Link to post
Share on other sites
16 minutes ago, Seraphim88 said:

Shouldn't an AI be able to learn proper grammar, use of context and whatnot? It might not be on the table right now for everyday consumers, but I feel like it's something that could really revolutionize translation processes in the near future.

There are already programs like DeepL that are very good at this, just not for Japanese to English, as far as I'm aware.

Share this post


Link to post
Share on other sites
6 hours ago, Seraphim88 said:

Shouldn't an AI be able to learn proper grammar, use of context and whatnot? It might not be on the table right now for everyday consumers, but I feel like it's something that could really revolutionize translation processes in the near future.

Yes, but we're not there yet, and current AI is not even really trying. I have no idea what you mean by 'near future' but a good translation AI will need to have a theory of mind - understanding how people think - which is honestly disturbingly close to artificial general intelligence, which has a non-negligible probability of killing us all if certain conditions prove to be true. Very optimistic (or pessimistic, if they think it's going to kill us all) AI researchers might think human-level AGI is about 30 years off; others think it'll be more like a hundred.

"Context" is really a big problem to figure out because context is a lot of data. Context means understanding that someone is likely to be miffed if you insult them, but context also means understanding that the author may have previously stated in script 1 of the 30 you've read in so far that your character is a robot that does not feel insults the same way as humans. Or for an example that actually happens in something I've translated, figuring out that the re-use of a verb about a certain relationship 300 lines prior is significant and adjusting the translation based on that. Context means understanding what was happening in the world when the author was writing the piece. Context means understanding culture.

These problems occur even on the micro scale with Japanese because understanding how people think is crucial to interpreting a conversation correctly, and you are going to have a hard time reducing this down to something simple enough for a computer to learn with the time, data, and memory we can give it at this stage or to be programmed to do.

Share this post


Link to post
Share on other sites
13 hours ago, kivandopulus said:

Atlas and LEC are non-neuro machine translation services, so it's unfair to compare them with Google translate. I show comparison with other neuro machine translation services like Bing and Yandex.

 

5 hours ago, ChaosRaven said:

I'm kinda suprised how bad Atlas seems to perform in comparison to LEC and sometimes even Google Translate. Maybe I should change to LEC - it seems way better than the other two. At least judging from the given examples.

Don't use Google Translate or similar statistically-driven translation algorithms to play visual novels.  You want to use a rules-based translator that translates predictably (even if it's wrong), not a guesslation.  Meta-knowledge--knowing what you know, and what you don't--is important when faced with unreliable information.  Statistical translations often remove the metadata--the messy grammar and such--that cue you in when the machine translator is messing up.  If you know the translation is wrong, you can use context to fill in the gaps.  If not, you'll just end up more confused.  Also, rules-based translators have predictable quirks that you'll eventually learn.

As for context, you're going to need it whether you're reading the Japanese or a machine translation.  The machine translation doesn't need to incorporate the context necessarily because there's always a human reading it.  What's important is that it doesn't assume the wrong context, or that the human can identify when the context is wrong and fill it in appropriately.

Edited by sanahtlig

Share this post


Link to post
Share on other sites
14 hours ago, Seraphim88 said:

Shouldn't an AI be able to learn proper grammar, use of context and whatnot? It might not be on the table right now for everyday consumers, but I feel like it's something that could really revolutionize translation processes in the near future.

Yes, but as a software engineer myself, I can tell you an AI capable of that is still decades away if not over a century...

Share this post


Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.


  • Recently Browsing   0 members

    No registered users viewing this page.

×
×
  • Create New...