Vocab Translations and Cloze (Update) May-29-2024

:bunprogold: Hi all! :bunprogold:

Today we have an announcement for everyone that is currently using the Bunpro vocab decks, and those that will be using them in the future.


:bomb: The Problem :bomb:

We have always had a system of manually translating and checking all of the content that goes on the site in order to focus on providing the best out of the box experience. This was the case for the translations for N5, N4, and N3 sentences in our vocab decks.

However, as our team of native speakers produce more and more Japanese content, we have been unable to keep up with providing English translations to match.

Unfortunately, this lack of English translations for many sentences is preventing us from being able to offer the Fill-in (Cloze) review type for much of our vocab.


:rice_ball: The Solution :rice_ball:

We have created a new ‘interim’ step for all vocab on the site that has Japanese sentences, that will allow reviewing of vocab using the Fill-in review type.

As can be seen here in these examples, they now have a set of { … } tags at the end, rather than a highlighted portion within the English sentence translation itself. If you see sentences with this sort of translation, what it means is the following:

  1. The translation comes directly from ChatGPT, and has not been checked by staff.

  2. The highlighted portion at the end is the ‘dictionary definition’ of the word, as seen on the Vocab page.

As you may be able to tell from these examples, the reason for why the definition is required at the end is due to the way that ChatGPT translates sentences.

In many cases, rather than using one of the ‘official’ translations, something much looser will be used. This is neither a good nor a bad thing, it is just the way language works, as many different translations can fit in many different circumstances. However, as most of us tend to gravitate toward one specific translation when learning, not seeing it in the English translation can make the target word quite hard to guess.

This change will allow us to make cloze review sentences available for any vocab that we have Japanese sentences for a lot faster than our old system. However, we do not want to keep this list of definitions at the end of sentences forever, which leads to the next new feature…


:vertical_traffic_light: Quality Checking :vertical_traffic_light:

We are currently working on a more streamlined system to allow for the reporting/suggesting of changes so that we can begin the process of improving the translations. In the image above, you may have noticed this symbol next to the N-level.

スクリーンショット 2024-05-28 111958

This symbol means that the English translation you are looking at has not been verified by staff (came directly from ChatGPT), so may contain inaccuracies.

All users will be able to see this. However, the users that have access to this new feedback system will have the additional option of being able to click on the icon and open a suggestion box where they can input their own interpretation, including highlighting on the most accurate word in the sentence.

This looks a bit like what we see in the next image:

In many cases, the ChatGPT translations are very good to begin with. So in a case like this, the user would simply need to highlight the word, make any changes they feel are warranted, and press submit.

After submission, Bunpro staff will be able to see it behind the scenes, and accept the new translation straight away if there are no issues, and make the new translation the default.

For sentences that have already been verified by staff and the users as ‘correct’, you will see this symbol rather than the translation icon.

スクリーンショット 2024-05-28 115250

Anyone who is interested in trialing this new feedback system is welcome to join. Please just send us a message. We will mainly be selecting people based on level, demonstrated understanding, and a positive history of involvement/helping out on the forums.

:exclamation:Benefits of this system:

  • By using ChatGPT, we will be able to create Fill-in sentences straight away for all our current un-translated Vocab sentences.
  • The use of symbols on sentences will show users straight away what the status of the translation is.
  • Having a transparent system like this will allow users to have a higher level of confidence in all translations, and get a better idea of which translations have been Bunpro/community verified.
  • We may be able to eventually use this system as a way to introduce extra/alternative translations, such as more literal interpretations.

Two final notes:

  1. Despite having manually checked and translated the N5, N4 and N3 vocab sentences on the site, we though it would be a good opportunity to allow those to be part of the trial, to see where they can be improved. These sentences do not have the { … } definition at the end.

  2. Our use of AI within Bunpro is strictly limited to tasks like this (providing reasonable translations or TTS for 40k+ sentences), and only as a stopgap. We do not, and never will, use ChatGPT to produce core Japanese learning content.

Please give us a bit of patience over the next few weeks as we make all of our remaining vocabulary with sentence examples available to study through Fill-in.

We hope that this new system will increase the speed at which we are able to produce new content, and will also allow us to facilitate a more effective feedback loop to continue to improve the content on the site.

We thank everyone for their ideas and feedback, and are looking forward to seeing how we can continue to build this new set of features into something that positively impacts all users’ learning experiences. I hope you are all having a fantastic day! :blush:

30 Likes

just out of curiosity what chatgpt model did you use (3.5/4/4 turbo/4o) if you know (its fine if you dont)?

5 Likes

4 for the vast majority. The quality difference between 3.5 and 4 was pretty stark.

Edit - We’ll pretty much always plan on using the latest model unless there are some exclusions beyond our control preventing us.

7 Likes

Thank you for making it clear when chatgpt is used. Is a great tool, but some learning apps use machine translation and don’t tell you, which can be misleading.

Im interested in seeing how it works!

5 Likes

Will this also work for using Reading and Reveal & Grade? :pray:

4 Likes

Yeahp!
Anywhere where there are sentences

5 Likes

Neat! As I’m just getting started on N3 content I’m still quite some time away from most of these auto-translated sentences, but I’d be happy to keep an eye out and submit feedback for any N5-N3 sentences whose translations or highlighting may be improved (Not that I think that I’d be better at translating Japanese than any of your staff, but I have definitely encountered some sub-optimal highlighting from time to time)

5 Likes

I’d like to participate! I sent @Asher a message.

Will be great to be able to use Fill-in/Cloze style for many more vocabs, even if the translations are not yet verified. And I certainly don’t mind providing feedback for tricky translations. :smiley:

3 Likes

Amazing! Thanks for this update! I’m interested in helping out with verification and will send you a DM :blush:

4 Likes

Is there any update on the progress of this availability?

I assumed that it included the N1 vocab, but I just came across 検事 (JLPT N1) | Bunpro, and it was still on Translate style, even though I had gone to the N1 deck and tried to update all my reviews to Fill-in style.

Then I checked the N1 Deck and I don’t think there are any vocabs that have been updated to Fill-in. And the individual vocabs do not allow you to change them manually either.

Maybe I was getting ahead of myself? I know that many/most/all(?) of the N1 vocabs do have sentences available with {…}. Is there some reason they cannot be used with Fill-in style yet?

[It seems that the N2s do allow for switching, BTW.]

2 Likes

Glancing at the N1 sentences, it looks like some still don’t have the target word highlighted in the example sentence, which I’m guessing means they can’t enable Cloze yet.

That being said, if all of the words now have example sentences and the meaning added at the end in brackets, would it be possible to enable the Reading / Reveal & Grade mode in the mean time?

1 Like

The purpose of this new Translation Suggestion feature (in the middle of the OP) is to allow approved users to add such highlights to these sentences in the first place. The best way to do that is to allow us to work on them as we come across them in reviews. N2s already work this way. Just need to enable the same feature for N1s and we can start working on those, too. Unless I’m missing something (?).

3 Likes

Hey sorry for the late reply. Have you been given access to the Translation Suggestion feature? If not, I would reach out to someone on the Bunpro team to get access. I’ve had it for awhile, and have been putting in suggestions as I’ve come across them. That being said, even though I have this feature enabled, I have been doing it through the main Vocab page when reading through the Example section, not the Learn/Review mode.

Perhaps @Asher or @veritas_nz could provide an update? I’m curious to hear what the timeline is as well, especially for Reading mode, which shouldn’t require highlighting and may be an intermediary step if Cloze can’t be enabled yet.

1 Like