Non-English speakers, how do you deal with Bunpro?

[pjorge]

I was a translator (English to Spanish) for many years, so “making sense of English sentences” was kind of my job and I still do most of my reading in English. I actually find Wanikani more complex because the meaning is usually given as a word you then have to put into context (or, worse, there is some kind of cultural reference hard to pin down). Most of the time on Bunpro I try to find the “Japanese logic” of the sentence, mainly “why do they say that like this?” So being used to two languages is helpful because I know there is not a correct way of expressing something.

Having Bunpro is another language would be great for many speakers of that language, like Spanish. There is a lot of great Japanese learning resources in English and I am quite lucky to be able to use them (I have lifetimes in both Bunpro and Wanikani because they are great). But I feel that creating a Bunpro version in say Spanish would be such an enormous undertaking that doing it wouldn’t be worth the effort.

So, my advice: find the “Japanese logic” in everything Japanese.

[dmitrii]

I am not native english speaker, only speaking skills. But Bunpro too hard for me actually, i use this for translate https://www.deepl.com/ and add to Notice

It doesnt work for Example, you can improve your English ¯_(ツ)_/¯

[Jihatto]

I highly disagree with anyone who stated that “English is enough” or “you should learn English first”. This is such a narrow world view. Of course, naturally, people who are on this forum can understand English and will tend to think like this. But the reality is so far from this!

80% of the worlds population doesn’t speak English. And even the 15% of the world population for whom English is a second language, would prefer to learn in their native language. In other words, you’re missing out on a lot of growth potential if you decide not to localize Bunpro to any other languages.

I can tell from my own experience, that there are many young people who are interested in learning Japanese, but they simply don’t do it because they’re already struggling with learning English in school and if they used a tool like Bunpro, they would need to learn two languages at once! Learning one language is already a daunting task, so nobody is gonna learn two at once!

Now, one could ask the question “which language to add first”? And in the past, I could have understood when the answer would be “it’s not worth it, adding a single language is a daunting task, too and there’s no single language large enough making this worth it”.

But things have changed with the introduction of AI services like ChatGPT and them having an easy-to-use API. Unlike machine translation services like Google Translate which existed before, these services can understand the context of things like translations or explanations pretty well and provide a good translation, that doesn’t just translate word for word. But you can also specify when translating with AI tools that they should try to find an as-close-as-possible translation to the Japanese original and adjust the explanation accordingly where a closer translation or explanation exists for the given language. This will work pretty well 95% of the time, and the other 5% will still be convenient for learners to have a translation that is closer to the English one, but still not English.

When I say this, I’m talking from experience. I developed a translation tool (for mobile apps) myself for years, and when I migrated the tool from machine translation services to AI tools, the improvement in the accuracy and the possibilities for culture-fitting the translations were huge!

In other words: I would argue that rather than hiring people for each language, invest into the development of internal tools that are optimized at translating Bunpro to other languages. Just don’t forget to document clearly on the website and upon registration in other languages, that you don’t provide official support for Bunpro in other languages than English and therefore can’t guarantee the correctness. That you’ve used AI to translate to these for the convenience of your users and that you recommend learning in English for the best experience.

If you do this right, e.g. by letting the AI proofread everything after initial translation, I promise that 98% of the translations will be correct in the most widely spoken languages, and that you will reach way more users (especially younger ones) all around the world. And you would also make many of your existing users happier, including me.

I would highly prefer learning Japanese from German myself to save my brain the extra thinking cycles needed to process everything in a second language. Every day, I have to look up one or two words during study simply because either the explanations or the sample sentences use words I just haven’t come across yet because I’m not a native. It’s frustrating and slows me down.

[Buntebunpou]

The problem with ChatGPT is, that it is not intelligent. It’s just a very advanced text parser and generator so to speak. That also means that it is really good if a topic was covered a lot in its training material (i.e. “the internet”). But it is an absolute shitshow if the material was not covered or only in a bad way.

I’m also German and I understand your problems very well (especially with Bunpro as German is a bit more similar to Japanese and going through English makes things worse than going from Japanese to German directly).

I therefore also used ChatGPT (model GPT-4o) a lot with Japanese. The results are mixed. Using it for translating a text is great, the best translator by far, much better than DeepL. Also using it as a talking partner to train speaking, works okay (but is pretty boring).

But using it for Japanese grammar where the training material is missing is HORRIBLE. It generates just useless responses and halluzinates massively. Unfortunately there is only very basic material about Japanese grammar in German on the web. So it can explain N5 and N4 stuff quite well. But beyond that it gets just unusable. I tried it many times but in the end it can only translate English material. And that doesn’t work for everything, especially the grammar points where the English translation is quite confusing for Germans.

It wouldn’t work for Bunpro.

Besides, it also just repeats the stuff that all grammar textbooks do in a bad way (e.g. calling ka a question particle, calling -i/-na adjectives, talking about a masu stem and so on). Because that was its training material. Shit in, shit out, that’s the principle of Large Language Models unfortunately.

Same for other stuff like programming. If you need a very specific piece of code and all training material that covered it, was really bad code, ChatGPT will also create bad code. It even adopts bad behaviors from incompetent programmers.

[Jihatto]

You are totally right to question LLMs like ChatGPT in general. But for language learning in particular, it is extremely well suited! It’s actually exactly what it was trained for, hence the name “large LANGUAGE model”. Its strongest point is to understand and produce language! And like I’ve said, I’m speaking from experience of using it professionally exactly in this area. It works well.

That’s not so say “you are wrong”. You have a point: If you asked ChatGPT to give you a grammar explanation of a random grammar point, it would fail to produce high-quality explanations or examples. Because, like you said, it was not trained on a lot of high-quality content.

BUT: Bunpro already has high-quality content in English! And like you said, it excels at translating a text, and that’s what is would mostly be doing. But if you put all the explanations of the grammar together with the example sentences into one prompt and tell it exactly how to behave during translation, it will come up with an awesome translation that doesn’t just translate word for word, but also adjusts things to apply nuances in the target language and make the meaning closer to the Japanese usage.

In other words, because you bring up programming as a bad example: I 100% had the same experience as you, when I ask an open question like “give me code that does X” it will produce outdated, low-quality code with a lot of “best practices” broken in it. Since, after all, it’s just a statistical model and doesn’t really “understand” what it’s doing.

BUT: When I give it a bunch of my code and tell it apply the same thing I did in my code in the new code, it’s transferring capability is quite awesome. It will save me a lot of time and the results will be 95% correct. And if I do a second iteration where I let it review what it did and compare it to what the goal was, it can even get to 98% accuracy. For the last 2% I need to fix things manually to get it just right. But for me as a user (and many others) 98% is “good enough”, especially if the 2% isn’t “completely wrong” but just “slightly off”.

So, I disagree and strongly believe that it would work very well for Bunpro!

[Jihatto]

To prove my point, I just let Claude (my preferred LLM) translate the explanations for the last grammar point I learned into German using this very simple prompt:

Can you please translate the following Japanese Grammar explanation provided in English into German? Make sure that the main points made remain the same and the Japanese example sentences are not changed. But feel free to adjust both the explanation and the example translation if there’s a closer translation or more similar grammatical point in German.

Here’s the English original, please give me the German variant:
<pasted text>

The result is this, which I think is a pretty accurate translation:

In unserer zweiten Lektion über と betrachten wir die häufige Übersetzung ‘mit’, die mit diesem Partikel verbunden ist. と zeigt eine Beziehung zwischen (A) und (B) an, unabhängig davon, ob diese Beziehung als ‘und’ oder ‘mit’ übersetzt wird.

In beiden Fällen führen die mit と markierten Nomen normalerweise eine gemeinsame Handlung aus (diese gemeinsame Handlung kann auch einfach das Existieren sein).

• カレーはご飯と食べるよ。
  Ich esse Curry mit Reis. (Ich esse Reis ‘und’ Curry) *
• コーヒーを妻と飲む。
  Ich trinke Kaffee mit meiner Frau. (Ich ‘und’ meine Frau trinken Kaffee) *

Wenn と als ‘mit’ verwendet wird, bedeutet dies, dass beide/alle Personen oder Objekte gleichmäßig an der Handlung beteiligt sind. Im Gegensatz dazu hebt に hervor, dass eine Person das Ziel einer Handlung ist, was zu einer eher einseitigen Bedeutung führt.

• 先生と話す。
  Ich werde mit meinem Lehrer sprechen. (Betonung liegt auf dem gegenseitigen Gespräch) *
• 先生に話す。
  Ich werde zu meinem Lehrer sprechen. (Betonung liegt nur auf einer sprechenden Person) *

Hier sehen wir, dass bei Verwendung von と beide Personen sprechen. Es ist ein gegenseitiges Gespräch. Wenn hingegen に verwendet wird, wird die mit に markierte Person ‘angesprochen’ und nimmt normalerweise nicht aktiv am Gespräch teil.

[casual]

@Jihatto

For the last 2% I need to fix things manually to get it just right

I think any company including Bunpro would probably want to have a native speaker of the target language (e.g. Spanish) on staff before putting out that localization to customers. I’ve worked on enough projects localized to several markets to notice that the result produced even by well-meaning humans from an external agency ends up quite awkward.

Doubly so if we are talking about teaching materials like Bunpro. Even if the translation is mostly accurate, how would one judge if it actually is helpful for learners without being a native speaker too?

For users who are happy with purely LLM translations of English that’s already on the site, aren’t there already browser addons and OS features to do that automatically? The approach of “instantly LLM translate anything on your screen” will only become more prevalent as all devices get dedicated co-processors. So is there really a lot of benefit to providing a built-in localization of the same LLM-only quality.

[Mikan28]

I am not an english speaker myself, I am french and even if my comprehension level is near-native (my writing is not), I sometimes struggle with the explanations in english, but at the same time, it is because there is a few really good japanese teaching materials in french and a lot in english that I get to ameliorate my level in english so I guess it is a win win situation ?

However, i think I would less struggle with some grammar points if the explanations were in french, because as @Buntebunpou said, it is hard to understand the slightest difference between two (or more) nuances of “even” or “should” etc

[Buntebunpou]

But you are missing my points here. Just translating the English text wouldn’t improve the parts where you would pick a different approach for German. And that was my main point why I started this thread back then.

Just a few examples:

• suru can be used to cover cost, we have a nearly identical structure in German: “Das macht 10€”
• I can’t come up with a good English translation of “goran” but German has one: “sich begeben”. That’s also a verb that is used nearly exclusively in very polite speech in German. It’s not exactly the same as “goran” but it would help to explain it for German learners.
• miru (German: sehen) can be added to verbs if you plan something, in German you can say “ich sehe zu, dass ich das mache” or “ich sehe, was sich machen lässt” or “mal sehen, ob man” which are quite similar to how Japanese uses miru in this context.

A good human translator would integrate something like that.

Also, your example is N5 level where there’s a lot of German learning material that helped training the LLM. And it also is something that is identical in German and English.

Try something complicated from N2 or pick a grammar item that exists in English and Japanese but does not exist in German like the present progressive.

I would guess that a LLM maybe gets translated 60% in a good way, no way it would be beyond 90%. And when it comes to all the example sentences Bunpro provides I bet a LLM would butcher a significant amount of them because it doesn’t get the context with the grammar point and just translates it which sometimes will kill what the example was for.

[adrianh]

For me, learning through strict translations and covering all of the specific cases where each word or expression have a slight different meaning in English (or any other language) seems impossible.

I take Bunpro translations as guides, but try to internalize the literal meaning they try to convey, it does not matter if the English sentence does not make much sense grammatically.

For example, this sentence:

私が会議に出る事になった。

Bunpro translation:
It was decided that I will attend the meeting.

How I try to see it:
It became a thing of me coming out (as in appearing) to the meeting.

I don’t think a language lacks the literal means to convey a meaning another language can express, we just have to find the human logic and get creative with it.

[EmperorLicht]

As a fellow German who only started learning English after finishing school, I now, at age 29, hold a C1 certificate from my vocational training and study Japanese at university. If your English is as good as you claim, I don’t see why you’re struggling. While some sentences in Bunpro’s practice exercises may sound a bit unnatural or unclear, I’ve realized that this is often due to my own lack of understanding of Japanese rather than any issue with the English itself. When the Japanese is clear, the English is usually clear too; though I have to admit, there are still cases where some ambiguity remains. That said, tools like DeepL or ChatGPT can easily help with that. For Japanese learning jpdb.io is also a great resource.

But also, when I hit a wall from overwhelming myself with too many reviews, I often just start over. It may not be the most time-efficient approach, but I’m in for the long run anyway.

[Jihatto]

@casual

aren’t there already browser addons and OS features to do that automatically

Yes, there are, and I’ve tried them, they do a pretty good job at localizing the texts on the page, but they break other things on the site! Such as the Furigana on hover for Kanji. But even if I developed such a plugin myself, still all users would need to set it up with the OpenAI API and would need to provide a good prompt for translating each page and would need to pay the translation. Therefore Bunpro could save all users a lot of money and time to set all this up. Also, as users, we don’t have access to the structural data as Bunpro has internally, so just from the text on the site some contextual information may get lost.

@Buntebunpou I just picked the “Goran” example you talked about to see what the same prompt from above does with it. Here’s the result:

てご覧 ist ein höflicher Sprachausdruck im Japanischen, der ähnlich wie なさい verwendet wird. Er wird üblicherweise als ‘versuche bitte zu (A)’ übersetzt, wobei (A) das Verb ist, das durch die Verbindungspartikel て mit ご覧 verbunden wird.

• この本を読んでごらん。この本を読んだら日本について色々と学べるよ。 Versuch mal, dieses Buch zu lesen. Wenn du es liest, wirst du verschiedene Dinge über Japan lernen. *
• 外を見てごらん。虹が出ているよ。 Schau mal nach draußen. Da ist ein Regenbogen. *

Obwohl es als ‘versuche bitte zu (A)’ übersetzt wird, ist die wörtliche Bedeutung von 覧 ‘ansehen’ oder ‘durchsehen’ (in Ruhe und ohne Zeitdruck betrachten). Daher kann man sich てごらん ähnlich wie てみる ‘versuchen’ vorstellen, aber mit der zusätzlichen Betonung von ‘probiere es aus, wann es dir passt’.

Achtung Da ご覧 zur höflichen Sprache gehört, kann es nicht für die eigenen Handlungen oder Dinge verwendet werden, die man selbst ‘versuchen’ wird. Es wird ausschließlich verwendet, um anderen Vorschläge zu machen, was sie ausprobieren sollten.

• これ食べてごらん。あなただったら絶対に好きだと思うよ。 Probier das mal. Wenn du es bist, wirst du es bestimmt mögen. *

Achtung Manchmal werden ごらん und なさい zusammen verwendet. Dies verstärkt die Aufforderung leicht und kann als sehr milde Anweisung wahrgenommen werden. Deshalb wird ごらんなさい hauptsächlich von Vorgesetzten/Älteren verwendet, wenn sie einen Untergebenen/Jüngeren bitten, etwas zu versuchen.

• 見てごらんなさい。こんな酷いことをネットにかかれたんだよ、酷くない？ Schau dir das mal an. Jemand hat so etwas Gemeines im Internet geschrieben. Findest du das nicht schrecklich?

The explanation it gave for what changes it made was this:

I’ve adapted the translation to include the German modal particle “mal” which often serves a similar function to てごらん in making suggestions softer and more casual. The German “probier mal” and “schau mal” are very natural equivalents that capture both the suggestion and the casual, unhurried nature of てごらん. I’ve also maintained the distinction between the basic form and the more formal usage with なさい by using appropriate German phrasing that conveys the same hierarchical nuances.

Seriously, this is pretty good. Definitely not 60%. Maybe not the level of adjustment you expected, but my prompt was very simple and plays safe, which is I think, what should be done on the first iteration. And this kind of translation would be suuuuuuuuuper useful for my cousins, for example, who are just learning English in school and who try to learn Japanese in their free time. But unfortunately, I can’t recommend Bunpro (or WaniKani, or any great English source) to them. The lack of AI translations is simply making these tools inaccessible to 80% of the world’s population. And like I said, it should be clearly stated these languages are not officially supported, but are just AI translations for convenience for users for whom this is enough. My cousins wouldn’t mind at all!

I’m not one of the 20%, I can access these sites because my English is good enough. But even I would prefer an AI-translated German text (with a button to see the English original when needed).

[Buntebunpou]

But this translation of goran makes it sound more casual in German (it even says that itself) while in Japanese it actually makes it more polite. That’s the opposite!

And the LLM most likely had the necessary information because there’s a lot of stuff about basic keigo on the web.

That’s my main problem with LLMs - they just don’t understand stuff. But they still sound very convincing and self-confident when producing bad output.
For a lot of tasks they are still great though and to do an initial translation of something like Bunpro that has to be corrected and improved by a human afterwards they would probably save quite some time. But you still need a native speaker who fixes the output of the LLM. And after my experience that is way more than 5%.

I actually tried using ChatGPT for Japanese grammar a lot. The results were mixed. It did great for some things and it did terrible for others. And you notice very quickly that it doesn’t have any intelligence but just processes information from the internet. If there’s no information or only low quality information, it is quite useless. It can’t draw own conclusions. Unfortunately Japanese grammar in German is such an area with not much information on the internet. As long as you can translate the English information 1:1 it will probably do fine. For the rest not so much.

I guess that would need a conceptually different AI approach than LLMs.

[Buntebunpou]

That’s the best approach in my opinion because it helps a lot with thinking in Japanese.

But unfortunately it won’t work with everything and won’t help with the problem that there are a lot of synonyms with slightly different nuances (e.g. koto and mono in your example).

In the end exposure to Japanese over time will solve it anyway but coming up with examples that fit your native language better, helps getting it in the first place after my experience. Especially with something like the many patterns that all translate to “even“ in English while there are more fitting translations in other languages.

[Jihatto]

@Buntebunpou I haven’t used LLMs for Grammar learning in particular, so maybe I’m wrong with 95% and it’s actually more at 85-90% on the first try. BUT: My prompt was super simple and did not guide the LLM very well. A good prompt would create a much better translation accommodating for common issues in the translation and reducing them significantly by telling the LLM what to do. Because, as you say, it doesn’t “understand” the task you’re giving it. It’s only producing the statistically best result for the task you’re giving it. So with a good prompt, you get way better results. And you need to iterate on the prompt like 20-30 times at least to get to the high accuracy I was talking about.

You made a good point about the politeness level being important specifically for translations from and to Japanese. So I asked the LLM to review and revise its translation with the following prompt, now also considering politeness levels:

Please review this translation and your changes to fit the language and any decisions you’ve made for the given grammar point. Make sure to always provide a proper related politeness level translation in the target language to create a similar politeness level “feeling” for the German learner compared to how a Japanese person would feel for the Japanese original. And give me the full corrected translation if any changes are needed.

And it’s answer was this:

I’ll review the translation with a specific focus on politeness levels:

1. Looking at the examples:

• この本を読んでごらん → “Versuch mal, dieses Buch zu lesen”

  • This is slightly too casual. てごらん is polite but friendly. A better translation might be “Möchtest du nicht dieses Buch lesen?” or “Lies doch mal dieses Buch”

• 外を見てごらん → “Schau mal nach draußen”

  • Similar issue. While “mal” helps soften it, てごらん has more of a caring, guiding tone.
  • Better: “Schau doch mal nach draußen”

• これ食べてごらん → “Probier das mal”

  • Again, too casual. てごらん implies caring guidance.
  • Better: “Probier das doch einmal”

• 見てごらんなさい → “Schau dir das mal an”

  • This needs more formality as it’s ごらんなさい
  • Better: “Schauen Sie sich das bitte einmal an”

Here’s the revised translation with these adjustments:

“”"
Über てごらん
てご覧 ist ein höflicher Sprachausdruck im Japanischen, der ähnlich wie なさい verwendet wird. Er wird üblicherweise als ‘würdest du bitte (A) versuchen’ übersetzt, wobei (A) das Verb ist, das durch die Verbindungspartikel て mit ご覧 verbunden wird.

• この本を読んでごらん。この本を読んだら日本について色々と学べるよ。
  Lies doch mal dieses Buch. Wenn du es liest, wirst du verschiedene Dinge über Japan lernen.

• 外を見てごらん。虹が出ているよ。
  Schau doch mal nach draußen. Da ist ein Regenbogen.

Obwohl es als ‘würdest du bitte (A) versuchen’ übersetzt wird, ist die wörtliche Bedeutung von 覧 ‘ansehen’ oder ‘durchsehen’ (in Ruhe und ohne Zeitdruck betrachten). Daher kann man sich てごらん ähnlich wie てみる ‘versuchen’ vorstellen, aber mit der zusätzlichen Betonung von ‘probiere es in deinem eigenen Tempo’.

Achtung
Da ご覧 zur höflichen Sprache gehört, kann es nicht für die eigenen Handlungen oder Dinge verwendet werden, die man selbst ‘versuchen’ wird. Es wird ausschließlich verwendet, um anderen freundliche Vorschläge zu machen, was sie ausprobieren sollten.

• これ食べてごらん。あなただったら絶対に好きだと思うよ。
  Probier das doch einmal. Wenn du es bist, wirst du es bestimmt mögen.

Achtung
Manchmal werden ごらん und なさい zusammen verwendet. Dies verstärkt die Aufforderung und wird als höfliche, aber bestimmte Anweisung wahrgenommen. Deshalb wird ごらんなさい hauptsächlich von Vorgesetzten/Älteren verwendet, wenn sie einen Untergebenen/Jüngeren darum bitten, etwas zu versuchen.

• 見てごらんなさい。こんな酷いことをネットにかかれたんだよ、酷くない？
  Schauen Sie sich das bitte einmal an. Jemand hat so etwas Gemeines im Internet geschrieben. Finden Sie das nicht schrecklich?
  “”"

Key changes made:

1. Added “doch” in examples to better capture the encouraging, friendly tone
2. Changed the basic translation pattern from “versuche bitte” to “würdest du bitte” to better reflect the polite but friendly nature
3. Added proper Sie-Form for ごらんなさい examples
4. Adjusted the explanation to emphasize the friendly, guiding nature of the expression
5. Modified “mal” to “einmal” in more formal contexts
6. Added “freundliche” to emphasize the friendly nature of suggestions

These changes better reflect the politeness levels and social dynamics present in the Japanese original, while maintaining natural German expression.

It found the politeness issue and corrected it. This is just the second iteration of the prompt, and we are getting closer to 95% already.

Don’t get me wrong: I’m not saying “any prompt will generate good results”. I’m saying that one full-time employee who knows how to use LLMs effectively and maybe knows Japanese + another non-English language, will be able to iterate on the prompt quickly and get to surprisingly good results. Good enough for anyone who doesn’t speak English to profit a lot from Bunpro!

Of course, the Bunpro team can still hire people to look over these translations if they wanted to make the support more “official”. I’m not saying “they should not do it”, all I’m saying is “why haven’t they done the initial translation part yet and provided it as an option, it’s so low effort”!