Jump to content

Quick opinion about DeepL by someone that reads in Japanese


Bredan

Recommended Posts


For the first time, I tried DeepL.

English not being my mother language, I first tried it with English-mother language (based on alphabet) : it was astonishing. I can't say it was perfect, but honestly I was impressed.

I then decided to try it with Japanese (one chapter of a LN to be precise) : and again, I was impressed. The result was far lower in quality than my first attempt, but nonetheless, it was probably the best MTL I've ever seen. Also, note that the chapter I used had only a few dialogues, maybe it would have been a different result with a VN having a lot of them.

Oh, well, but everything I just said is quite well known I guess... then why do I post this ?

It's simple : I just wanted to warn the people that use DeepL, be it for VN or anything else. There's a deep flaw going with it : it is too good in a certain sense.

Too good ? WTH ? Well, DeepL seems to be able to give sense to a sentence most of the time. Yeah, give sense even if the translation fails (partially or completely), and you the reader might not notice because the sentence might even make some sense in the context, just that it has not the sense it is supposed to. While the Google Translation lot would most likely just send you an unintelligible text that would at least warn you of the failure.

I'm not sure there's a workaround, but I guess one should try, if possible, to use at least two engines in parallel and hoping to get the better of them.

Note : my level at reading Japanese is quite high when we are speaking of reading VN/LN/Mangas etc. I'm positive I encounter very few passages I'm not correctly understanding and, when this happens, I generally know enough to get that I missed something, allowing me to do a bit of research if I feel like it. I'm not sure how I would fare with true literature for example (a lot less good I presume, but I guess it would depend of the writing). Also, when possible I'm still using a text hooker to help me seek out the few kanjis or words I might not know or remember.

Link to comment
Share on other sites

I have used DeepL for a few visuals and some it gives a pretty good translation for, others not that great. I can't read Japanese, but I can understand some spoken. I am guessing the TL quality is due to the software at times, not having the right Hook, and other times because of how the writing is in the game. My main complaint with it is how bad it is with pronouns based on sex of the character referred to or just who it is talking about. Sometimes a lot of mental editing is required. However, I am thankful for DeepL as without it I would not have been able to read the masterpiece that is Sakura Moyu.

Link to comment
Share on other sites

1 hour ago, Kirashi said:

Wasn't deepl unusable for playing because it only translate a few words a day?

Hmm, I don't know about this. That was the first time I used it and I encountered no limitation other than not being able to translate the whole chapter in one go (generally, I translated paragraph after paragraph, or few lines if paragraph was too long). Maybe one chapter was not enough to get to the limitation you are speaking of ? Can't say.

Also, I can't guarantee it, but it seems people successfully use Textractor with DeepL plugin.

Edited by Bredan
Link to comment
Share on other sites

11 hours ago, Seraphim said:

There's a limit of 5000 characters a day. If you want more than that, you're going to have to pay their monthly rate of €8.99 (or €5.99 if you sign up for a year). If you use it a lot, it seems like a pretty decent price.

I used it on sugoi translator and never ran into a limit. And that us with playing visual novels all day.

Link to comment
Share on other sites

Seems like the limit is a bit unclear.
A 5000 character limit was listed on their website, but I just noticed another note on a different part of it, saying "With the DeepL API Free plan, you can translate up to 500,000 characters per month for free."

So maybe it's 500k a month and you can also only use 5k characters within a certain amount of time (which doesn't seem to be specified), so if you break either of those limits, it won't work anymore until the timer resets.

Edited by Seraphim
Link to comment
Share on other sites

As far as I'm aware there is a limit for 5k characters at a time. If you want to translate a longer text you need to feed it to DeepL in portions. Which might have a negative effect on translation quality as the AI works better the more context it has to work with.

Link to comment
Share on other sites

2 hours ago, alpacaman said:

As far as I'm aware there is a limit for 5k characters at a time. If you want to translate a longer text you need to feed it to DeepL in portions. Which might have a negative effect on translation quality as the AI works better the more context it has to work with.

I indeed couldn't translate the whole chapter in one go and did it in portions.

Link to comment
Share on other sites

  • 1 month later...
On 7/13/2022 at 11:55 PM, Kirashi said:

Wasn't deepl unusable for playing because it only translate a few words a day?

No, it translates everything. A couple of years ago, I'm thinking because it ran on previous-gen hardware, some words were not recognised. I still think this is the case, but the numbers have been reduced. So, it's not that the majority of it is not TLed, it's just some bits. Now, even names used in conversation would be TLed too, so "Sango" is "coral".

The quality of MTL is getting sluggishly improved, but the problem is more about grammar, sarcasm, double entendres, nuances, etc. VNs tend to be full of innuendos (if not keigo), so it's hard for it to keep track.

Link to comment
Share on other sites

The free version of DeepL can be used in browser and on PC and it does 5000 Character long blocks at once in the free version. If you want to use deepL with the api then you will get blocked after a while unless you run a particular github script that breaks this limit. There have been whole playthroughs done with DeepL after ppl were able to skip that limit and it basically allows you to translate games to a 30-40% accuracy if I can judge from the posts made by frequent MTL-playthroughers such as Jazzmonkay and GalacticLightN.

 

Link to comment
Share on other sites

I haven't used DeepL for playing VNs, but for making them:

I used DeepL to make a first run of translations for my game (The Last Secret) that then has been improved by human editors. I' ve written a Python program that makes an auto-translation of the Ren'py translation files in the language of choice using DeepL (with license), so that this step works fairly fast now. 

I've started with German as base language, since it has more information on genders and politeness forms than English. The translation into English was usually spot on, and merely required polishing. The polish still improved the game a lot. The translation into Spanish and Chinese needs a bit more time to take care of well, so at release we won't be done completely with them. (In Spanish, there's also the additional difficulty that gender differentiation in language is at times even higher than in German requiring if-constructions in the translation files.) For Japanese and Ukrainian we are still looking for volunteers to help with this, so not much experience there yet. Unfortunately, I cannot read Japanese, so I can't judge how good that translation is. I guess similar to Chinese which is most of the times understandable, but has mistakes.

Link to comment
Share on other sites

I tried Sugoi Translator (Offline) recently and was impressed by the strides made recently in machine-learning enhanced JP -> ENG translation.  My friend and I independently checked the translation accuracy of the latest offline translation model (3.3) in two different titles.  We agreed that accuracy has reached the point that those with some JP knowledge can use it as an alternative to dictionary lookups or for speed reading.  Subject/object confusion is still a major problem and therefore we aren't quite at the point where it's safe/recommended for those with no JP knowledge.

I ran some comparisons with DeepL and found Sugoi Translator (3.3 model) to be comparable to DeepL for VNs.  Sugoi Translator has several significant advantages over DeepL:

  • Can be run on the host computer with low translation latency (<1s) when properly configured.
  • Optimized for VNs, and can recognize and translate odd speech patterns used by anime characters.
  • Freeware with no rate limits.
  • No online connection required.
  • Doesn't spontaneously break all the time because you were blacklisted or the server API changed.

The tool however is non-trivial to properly configure and will require some persistence and resourcefulness to fully optimize (e.g., update the model, use CUDA, and seamlessly incorporate into a workflow).  Documentation is scattered and updates are frequent and not automatic, and therefore many users are probably using a non-optimized version.

I'm sufficiently satisfied with Sugoi Translator that it's replaced ATLAS in my workflow.  Until now I've found statistical machine translation alternatives either too unreliable (Google Translate) or too unwieldy (DeepL), but Sugoi finally convinced me that real progress has been made.

unknown.png

unknown.png

 

Edited by sanahtlig
Link to comment
Share on other sites

  • 4 weeks later...
  • 3 weeks later...

A 'bit out of context but still connected' quick update : have you followed the last few years progression in the department of AI ? I didn't until recently... am I reading an Asimov novel ? Are we really already debating the 'conscience' of this or that AI ? Of course, such an AI still needs one of the costliest and big 'datacenter' in the world. But still.

I envisioned MTL everywhere in the near future, but not that near... by the way, time to debate : should we still call that MTL, since some of those AIs define themselves as human beings ?

Seriously, do we need to invite Sarah to diner next weekend ?

Edited by Bredan
Link to comment
Share on other sites

I'm playing Kuro no Kiseki 2 (Jrpg) on my PS4 with a combination of Capture Card + VN OCR (using DeepL as translation engine) and the result was better than expected. I understand some japanese , so it helps a lot when the OCR get's wrong Kanjis, and ofc the context and pronouns.

I think it's a great tool if you're interested in learning the languaje. Wouldn't recommend it otherwise.

Link to comment
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...