It is natural for people who do AI self-publishing to read cutting-edge academic papers, follow remedial courses, and improve their knowledge system.
Just to keep up with the times, one has to fight tooth and nail not to fall behind.
But right, I am not afraid of your jokes, my English is really bad, the most skillful shout out of that English is:
I'm fine, thank you, and you?
Truly, in reading the various documents, even though the translation has been augmented with the Great Model, it still suffers from all sorts of tortures.
Like yesterday when Beanbag posted a paper on AI images.
I sought to go through their previous SeedEdit paper and learn against it.
This is it.
I don't know much about other papers, but these kinds of AI papers are generally contextually logical to the point of death, and filled with a lot of highly specialized terminology and academic expressions.
There are a lot of terms like Out-of-Domain (OOD) and DiT architecture.
Just the specialized academic terminology part, it's not bad. But when it comes to some specialized terminology with complex expressions, I'll be honest, I usually use immersive translation with OpenAI or DeepL's API now, and it's still confusing to translate.
A lot of times, it has a hard flip for specialized terminology, even if you don't understand it. For a non-specialist like me, it just looks like a real struggle.
It just so happens that I recently watched the Tamabuchi Tanten do another book on the subject ofAI TranslationThe quality of the curation is pretty awesome. So I also thought, why don't I review it myself.
Let's see which big model, in this scenario, has the best translation quality.
Say it and do it.
So, my partner and I spent 2 days livering and testing N number of questions, and found what we think, is the best big model for translation at the moment.
Directly to the conclusion: Netease Yudao, an old translation vendor out of the AI big model Ziyi Translation 2.0, actually hanged everything.
Tell us how we measured the translation review piece of the big model.
The dimensions of the test were also simple, just 2 points sorted out from my usual tortured experience.
These AIs must:
1. speak human; 2. don't turn over without a job.
The first point is to speak the language of the people. It is good to explain that the content of the translation is in line with the language habits of Chinese people.
The second point, don't flip if you don't have a job, is even better. Technical terms and some proper nouns, if you do not know how to turn can not turn, but do not turn hard. Other content is basically correct on the line.
In terms of runners, I went with the standard foursome of Grok3, ChatGPT-4o, Claude 3.7, and Gemni 2 from overseas. Domestically I ran all the ones I could as well, but due to space constraints, I'll just show the DeepSeek R1, Smart Spectrum GLM-4, Aristotle ZiYi Translator 2.0, and Qwen 2.5-Max as the quartet inside the article.
DeepSeek R1 is the only inference model thatThough theoretically I shouldn't use inference models to measure translations either, because it's not practical. But who asked for people who are just too hot = ==
I'm giving these big models outfirst question, is a slightly more complex paragraph in this SeedEdit paper that involves a bunch of high-frequency jargon and also contains academic citation information, author names.
Let's look at domestic production first.
speak humanelyaspect.
Both Qwen and Wise Spectrum GLM need to work on their human-speaking, and the most iconic sentence is this one:
Qwen: We note that our approach focuses primarily on application scenarios in the HQ-Edit benchmarks where we want to modify T2I-generated images based on arbitrary instructions.
Smart Spectrum GLM: We note that our approach focuses on application scenarios in the HQ-Edit benchmark where we wish to modify T2I-generated images with arbitrary commands.
The ability to have three of the same subject in a single sentence makes it very difficult to read and breaks it strangely.
ZiZi Translator 2.0 does a great job in reverse. It has less subject abuse and avoids a bunch of overly repetitive pronouns. Especially when dealing with sentences like "that" and "where", it does a good job.
DeepSeek is slightly weaker than Ziyi Translator 2.0 in terms of human speech output, and the translation cavity is a bit heavier in mechanical sense. An obvious example is the first sentence, which is not concise enough and the subject disappears.
Don't turn it over if you're not working.In terms of that, both Ziyi Translator 2.0 and DeepSeek are significantly better.
rightOut-of-Domain (OOD) teston the processing, only Qwen translates to cross-domain testing. Wisdom Spectrum GLM and Qwen have a good understanding of thereal in-the-wild image inputsThe translations are a bit hard to top. In fact, just flip it directly into an image input of a real scene.
The foreign group, in the English translation of the performance than the domestic group to be slightly worse, that the feeling of English Chinese really lingers, read very not fluent.
The best performer here is ChatGPT-4o. Its translation is overall clear and accurate, and does a good job of presenting the technical content and academic tone of the original text.
Gemini is basically in the same league as Claude, with almost as much human output.
Grok 3 does okay, but it's the same old thing: crazy subject stacking.
This first question is just a "normal" paragraph in an academic paper, and it is not too difficult to understand the semantics.
But when you look at hard-core technical papers, I don't know if you have the same feeling as I do: the AI may perform OK when confronted with content in the main text. But when the object changes to the content of the figure notes, there is a high probability that it starts to jerk around.
Since there aren't any complex paragraphs in the figure notes section of this SeedEdit paper, I've taken a paragraph from the classic sacred text Attention is all you need as aSecond major question.
This time, it's clear how well the individual AI macromodels are able to translate English in the face of such implicit contextual constraints.
The first long difficult sentence at the beginning is really a summary of the picture, however concise.
Wisdom Spectrum GLM's understanding of long difficult sentences in the context of diagrammatic notes is a bit off.
DeepSeek and Qwen chose to make a break in the middle of the sentence, which is not meritorious. However, at the end, they both chose to "recommend viewing in color", which is a bit more casual in an academic context. But it's not a big problem.
Ziyi Translation 2.0 is very concise in its presentation of a sentence, very much in line with the norms of a normal academic graphical note. However, there is a small problem. Because heads This word had appeared in the previous sentence, and Ziyi Translation 2.0 directly omitted it in the next sentence, giving the translation asDifferent colors represent different heads.This can actually put some pressure on understanding.
Gemini has the same faults as GLM, ChatGPT-4o directly hard translates following in the first sentence to tracking, Pass.
Claude's translation of the first sentence, which is also awkward as hell.6 layers xxx in the 5th layer.No one talks like that.
In addition to the two test questions above where the object is plain text, another point that particularly complicates my mood when reading papers on a daily basis is when a bunch of math symbols are inserted into the text. Translators are easily confused.
So I'm also going to take that as the third big question and test how these eight models deal with these number symbols.
The text is still using an embedded math equation from Attention Is All You Need.
Overall, each of the big AI models actually performs quite well in terms of understanding embedded mathematical formulas.
Although the mathematical notation format of d² is typographically less than regular in the treatment of some models, there is nothing confusing about the logical presentation.
DeepSeek and Ziyi Translation 2.0 are able to balance terminological accuracy with speaking human expressions when dealing with technical content, which clearly outperforms several other models.
Another obvious point I noticed while doing this test was the last sentence.
xxx, the approach we take in our model.
Obviously, this sentence should not be distinguished from the previous one.
Because in the Chinese context, it is actually a supplemental description of the foregoing. But all the seven big models except Ziyi Translation 2.0 give isolated translations.
A comma that gives all 7 big models a snub.
A set of process all measured down, only to say the Chinese and English translation, you will clearly feel in thespeak humanelyAs for the aspect, there is a way to this Ziyi Translation 2.0, is really let me unexpectedly feel the bull.
existDon't turn it over if you're not working.On, DeepSeek R1, Ziyi Translation 2.0, ChatGPT-4o is basically a echelon, terminology know to keep, not blindly play.
The most surprising point, at least in my own tests, is still that Yodo's Ziyi Translation 2.0 is the most effective, a point that actually rather surprised me.
But at the same time it does validate a "consensus".
In vertical scenarios, these industry giants that have accumulated data for N years hold resources and data in their hands that are really not rivaled by some generalized big model companies.
As I was figuring out what I felt was the best big model for translation at the moment, and getting ready to let it come in later to help me read papers and read technical articles, I realized.
Immersive translation did not leave an API interface for Ziyi Translation 2.0, it can't be connected to it.
And have a road to their own products, although doing a very good job, but for people like me every day in a variety of English sites to bounce back and forth also brush X and in the arXiv pickpocket the paper, the experience of the path does have a little long.
This one gave me a whole won't....
But if you're used to using apps, or if you're a cell phone mainstay, then I'd brainlessly recommend that you go inside of Yodo their own product.
For example, theirYoutube translationand Aristotle.
Screenshot translation, PDF translation and so on are all almost, simultaneous interpretation and so on are all there, this inside the AI translation is Ziyi translation 2.0.
But the most unfortunate thing is that there is no browser translation plugin that I want.
So, after a minute of hesitation on my part, it's easier to not have to go back and forth to multiple pages and save time to do my science reading in peace.
I still made a very silly der decision.
Hand rub a thesis-specific translator plugin that can pick up the Ziyi Translator 2.0 API by yourself with AI.
EMMMMMMMMM.
The plugin translates like this.
The user experience effect of this translation plugin is definitely not as good as those commercial plugins, but barring that, it's a no-brainer. Translation quality, more important than experience.
The old rules, this plugin I also put in the public backstage, everyone feel free to take, direct public backstage private message me "fy"Then the system will send you the file automatically.
Plugin installation and API access are actually exceptionally simple.
Let's start with how the API is obtained.
Go directly into the AI Open Platform at Aristo below to request an API:
https://ai.youdao.com/modelTranslation.s
In the top navigation bar, selectProducts & ServicesThere's one in there.Large Model TranslationJust click on it.
optionUse Now.
In the left navigation bar selectnatural language translation servicesinnerLarge Model TranslationClickCreating Applications.
The app name can be whatever you want, I wrote:PDF Translation, select the service field to be selected:Large Model Translation, access method selection:API, the application category is directly selected:Utilities.
After selecting these four items, just tap Confirm.
These are actually your app IDs.and the secret key.
On the Youtube API platform, as soon as you sign up with your real name, you'll get a $50 experience.
Basically enough to use for a while, probably high-quality processing of three or five thousand pages of PDF.
As for the installation of my hand rubbed Thesis Translator plugin it was easy.
Just follow the instruction chart below, step by step.
The first step is to unzip the translation plugin.zip I gave you.
Then open Google Chrome and tap the three dot button on the top right to go to Manage Extensions.
Drag the file in with one click and it installs automatically.
Open the Extensions button and put [Kazik Thesis Translator] on top.
Click this Settings button.
Enter the API key and API Secret that you got from registering on the Youtube API cloud platform, and tap Save Settings to use them.
After installing the plugin, you refresh arXiv and you'll find a [View Chinese Translated Version Online] button on the right.
Tap in to start the pure enjoyable version of the paper translation. You can click to hide the original text, directly is a Ziyi translated Chinese version. You can also, with one click, change the layout of the original translation layout, like the following.
Supported, too, by the graphic.
Some web pages are fine.
The whole process, literally, takes a few minutes.
The next step is to be able to navigate the world of papers translated from Chinese to English.
Honestly, this little hand rubbed plugin of mine is pretty crappy, so if Yodo or Immersive Translations will see this post of mine, I hope:
Either Yodo comes out with a plugin or Immersive Translation supports Yodo's API.
Really, I think browser translation is a scenario that is still pretty much in demand.
Of course, if you think the plug-in is too runt (and indeed quite frustrating), then you can also download the PDF and throw it into the translation of the direct translation.
Or just use their screenshot translation.
However, if you're like me these days, and you just like to slog through English literature on your browser, and you're also looking for translation quality, then I think it's really a good idea to consider scraping by and using my little crappy plugin.
May we all go further and further into the world of AI and translation and never have to scratch our heads and cringe at the language threshold again.
May you and I go farther on our AI journey.
Common Encouragement.
