AI translations of technical papers works well

This is somewhat off topic, but AI translations of technical papers is getting better every month. It is amazing how good it can be. If you are not a native speaker of English, you might consider writing in your native language and having an AI translate the paper. It will make a good first draft. You will need to make corrections and patch up the result, but it will save a lot of work.

I tried several programs and found this one works best for technical translations:

DeepL Translate: The world's most accurate translator

Translate texts & full document files instantly. Accurate translations for individuals and Teams. Millions translate with DeepL every day.

www.deepl.com

I recently had to translate a tedious legal document into Japanese. It is 22,000 words. I don't know much about the legal issues. It would have been a pain in the butt doing this manually. I ran it through this DeepL. Then I had my screen reader TextAloud read the English while I followed the Japanese, and vice versa. It was easy to find and fix the errors. You soon learn what kind of errors the AI makes.

(I recommend TextAloud for translations. It reads documents in many different languages.)

Here is a sample DeepL translation. These are two paragraphs from a paper by Storms.

The experimental literature describing the cold fusion phenomenon is reviewed. The number and variety of careful experimental measurements of heat, tritium, neutron, and helium production strongly support the occurrence of nuclear reactions in a metal lattice near room temperature as proposed by Pons and Fleischmann and independently by Jones. . . .

At the present time, understanding of the field is divided between those who do not believe there is sufficient information to require a re-evaluation of conventional beliefs about atomic fusion and those who believe that the data, although incomplete, strongly suggest the possibility of unusual nuclear reactions in a metal lattice. An approach that would do justice to both viewpoints is difficult to contain in one review. Therefore, I adopt the more probable viewpoint by assuming that a real phenomenon has been discovered and try to show patterns of behavior that would give some understanding of its characteristics. In a few cases, I attempt to show how the proposed errors have been reduced or eliminated and how various conventional explanations do not apply. Hopefully, those who do not believe that any one study is sufficiently complete to change their minds might be impressed by the consistent patterns of behavior among various studies and the considerable background information that is now available. This review is not intended as a substitute for reading the original papers.

DeepL translation into French:

La littérature expérimentale décrivant le phénomène de la fusion froide est passée en revue. Le nombre et la variété des mesures expérimentales minutieuses de la chaleur, du tritium, des neutrons et de la production d'hélium confirment fortement l'existence de réactions nucléaires dans un réseau métallique proche de la température ambiante, comme l'ont proposé Pons et Fleischmann et, indépendamment, Jones. . . .

À l'heure actuelle, la compréhension du domaine est divisée entre ceux qui pensent qu'il n'y a pas suffisamment d'informations pour exiger une réévaluation des croyances conventionnelles sur la fusion atomique et ceux qui pensent que les données, bien qu'incomplètes, suggèrent fortement la possibilité de réactions nucléaires inhabituelles dans un réseau métallique. Une approche qui rendrait justice aux deux points de vue est difficile à contenir dans une seule revue. C'est pourquoi j'adopte le point de vue le plus probable en supposant qu'un phénomène réel a été découvert et en essayant de montrer des modèles de comportement qui permettraient de comprendre ses caractéristiques. Dans quelques cas, j'essaie de montrer comment les erreurs proposées ont été réduites ou éliminées et comment les diverses explications conventionnelles ne s'appliquent pas. J'espère que ceux qui ne croient pas qu'une étude soit suffisamment complète pour les faire changer d'avis seront impressionnés par les modèles de comportement cohérents entre les différentes études et par les informations de base considérables qui sont maintenant disponibles. Cette analyse n'est pas destinée à remplacer la lecture des documents originaux.

The French version translated back into English:

The experimental literature describing the phenomenon of cold fusion is reviewed. The number and variety of careful experimental measurements of heat, tritium, neutrons and helium production strongly confirm the existence of nuclear reactions in a metal lattice close to room temperature, as proposed by Pons and Fleischmann and, independently, Jones. . . .

At present, understanding of the field is divided between those who believe that there is insufficient information to require a reassessment of conventional beliefs about atomic fusion and those who believe that the data, although incomplete, strongly suggest the possibility of unusual nuclear reactions in a metal lattice. An approach that would do justice to both points of view is difficult to contain in a single review. For this reason, I adopt the most likely viewpoint by assuming that a real phenomenon has been discovered and attempting to show patterns of behavior that would enable us to understand its characteristics. In some cases, I try to show how proposed errors have been reduced or eliminated, and how the various conventional explanations do not apply. I hope that those who do not believe that a study is sufficiently complete to change their minds will be impressed by the consistent patterns of behavior between the various studies and by the considerable background information now available. This analysis is not intended to replace reading the original documents.

English to Japanese and back to English:

The experimental literature describing cold fusion phenomena is reviewed. The number and variety of careful experimental measurements of heat, tritium, neutron, and helium production strongly support the occurrence of nuclear reactions in metal lattices near room temperature, as proposed by Pons and Fleischmann and independently by Jones. . . .

Current understanding of the field is divided between those who do not believe that there is sufficient information to warrant a reevaluation of conventional beliefs about atomic fusion and those who believe that the data, although incomplete, strongly suggest the possibility of anomalous nuclear reactions in metal lattices. It is difficult to include in a single review an approach that would legitimately evaluate both views. I have therefore adopted the more likely viewpoint and attempted to show patterns of behavior that would allow some understanding of the characteristics of the actual phenomenon, assuming that it was discovered. In some cases, I have attempted to show how the proposed errors have been reduced or eliminated and how the various conventional explanations do not apply. Hopefully, those who think that no single study is complete enough to change their minds will be impressed by the consistent patterns of behavior in the various studies and the considerable background information now available. This review is not intended to be a substitute for reading the original paper.

Some observations

A human translator can easily spot the problems and the differences between these translations. He or she will know where the problems originated, and will be on the lookout for them. As I said, you soon learn what kind of errors the AI makes. Here is a simple example. The last sentence in the original:

This review is not intended as a substitute for reading the original papers.

French:

Cette analyse n'est pas destinée à remplacer la lecture des documents originaux.

Back to English. With no important differences:

This analysis is not intended to replace reading the original documents.

Japanese back to English:

This review is not intended to be a substitute for reading the original paper.

It says "paper" instead of "papers" or "literature." That is because Japanese grammar does not have person or number. There is no distinction between the singular "paper" and plural "papers." So the distinction is lost going into Japanese, and does not make it back again into English. Any translator knows this. We keep an eye out for it.

No "person" means there is no distinction between "I ate lunch" and "you ate lunch." You can tell who ate by context . . . if it matters. The translator has to use his knowledge of the real world to decide who ate lunch. To disambiguate, as they say in the linguistics biz. Since an AI has no knowledge of the real world, it has difficulty doing this.

You might think Japanese people are often confused about who did what to whom, but you would be wrong. Then again, the lack of person and number sometimes makes for hilarious comedy routines. Especially with regard to sex. I leave the details to your imagination.

Quote from Gennadiy Tarassenko

How do I find this program to translate into English, please tell me how to find II.

Click on the box above that says "DeepL." It is a website.

DeepL dot com

It allows short translations for free. For long documents, you have to pay every month.

Quote from Curbina

I can’t tell you how much laughter this sentence caused me, and how much more laughter the next sentence provoked by confirming my original thoughts...

One of the funniest things I ever heard was a 19th century “rakugo” comedy routine about this. These routines are often in the form of a conversation between protagonists. Two thieves robbing persimmons. A husband planning a night on the town with his buddies. In this case, two neighborhood women are gossiping. One asks, “so how is everyone in your family? How is your grandpa doing?”

“Oh, he’s fine. Healthy and full of beans as ever.”

They talk about other things for a while. The second woman starts talking about the family cat. The first woman does not notice she has changed the subject . . .

This does not translate. Okay, reset and imagine the woman asked “how are your grandparents?” So, to continue, the second woman, now talking about the cat, says:

"You know how they are, though. They go into heat, and get all frisky."

"Your . . . grandparents?!?"

(Second woman misses the remark) "Yeah, you have to keep ‘em tied up in the kitchen. Or they will wander around the neighborhood. The next thing you know, you have offspring and who knows where they came from . . ."

And on, and on, more and more confused.

Quote from Alan Smith

He told me he sheds a few tears sometimes when thinking about all the undergrads he has taught having to abandon the profession to robots.

Even though I am technophile, I must admit that I have mixed feelings about it. I have been hoisted by my own petard. Or in modern terms, I voted for the leopards eat your face party, but I never thought they would come after me!

Still, the world needs easy translation, on a scale the human translators could never meet.

Quote from Alan Smith

He's glad he stuck mostly with poetry, where sometimes the content is so riddled with cultural references that an AI would also need to be a cultural anthropologist.

The other day I ran some modern poetry through AI, from English into Japanese. I was curious to see if it works. DeepL works best with technical and legal documents, but it did a poor job with poetry. Google translate was too literal. ChatGPT did the best job. The first version was too formal and not very lyrical. I suggested changes in the grammar making it less formal. More personal. She agreed that was a good idea and instantly made the change. I suggested a few wording changes and she agreed they were more lyrical and closer to the English in that context. After some iterations it came out rather well.

One of her comments regarding a change I suggested:

Yes, "あなたに告げる ["I say to you" -- old expression]" would also be a suitable translation for "I tell you" in the poem. It adds a sense of solemnity and clarity to the statement.

"Solemnity" is right.

I had ChatGPT translate Wilfred Owen's poem "The Parable of the Old Man and the Young." It came out quite well, I guess because that poem in simple language, more like a newspaper report than a subtle modern poem.

I ran a part of a 1908 novel from Japanese into English with ChatGPT. It was passable.

Paragraphs translated from the novel "Sanshiro" by Natsume Soseki

Translated by Jay Rubin:

When the train reached Toyohashi the sleeping man bolted up and left the car, rubbing his eyes. Amazing how he could wake himself at the right time like that, thought Sanshirō. Concerned lest the man, still dazed with sleep, had alighted at the wrong station, Sanshirō watched him from the train window. But no, he passed through the ticket gate without incident and went off like anyone in full possession of his faculties. Reassured, Sanshirō changed to the seat opposite. Now he was sitting next to the man with the mustache. The man moved across to Sanshirō’s former seat. He poked his head out the window and bought some peaches.

When he was seated next to Sanshirō again, he placed the fruit between them and said, “Please help yourself.”

Sanshirō thanked him and ate a peach. The man seemed to enjoy them very much. He ate several with great abandon and urged Sanshirō to eat more. Sanshirō ate another one. The eating continued, and soon the two of them were talking like old friends.

Translated by ChatGPT:

When the train arrived at Toyohashi, a man who had been sleeping slowly woke up and rubbed his eyes before getting off. I marveled at how well he could wake up so promptly. I thought that perhaps he had groggily mistaken the platform, but as I watched from the window, it became clear that wasn't the case. He safely passed through the ticket gate and exited like a sane person. Sanshiro breathed a sigh of relief and moved to the seat on the other side. Now he was sitting next to a man with a beard. The bearded man took turns sticking his head out of the window to buy peaches.

Eventually, he placed some fruit between the two of them and said, "Won't you have some?"

Sanshiro thanked him and ate one. The bearded man seemed to enjoy it and ate without restraint. He urged Sanshiro to eat more. Sanshiro ate another one. As the two of them ate the peaches, they became quite friendly and started various conversations.

Original Japanese text:

汽車が豊橋とよはしへ着いた時、寝ていた男がむっくり起きて目をこすりながら降りて行った。よくあんなにつごうよく目をさますことができるものだと思った。ことによると寝ぼけて停車場を間違えたんだろうと気づかいながら、窓からながめていると、けっしてそうでない。無事に改札場を通過して、正気しょうきの人間のように出て行った。三四郎は安心して席を向こう側へ移した。これで髭のある人と隣り合わせになった。髭のある人は入れ代って、窓から首を出して、水蜜桃[すいみつとう]を買っている。

やがて二人のあいだに果物くだものを置いて、「食べませんか」と言った。

三四郎は礼を言って、一つ食べた。髭のある人は好きとみえて、むやみに食べた。三四郎にもっと食べろと言う。三四郎はまた一つ食べた。二人が水蜜桃を食べているうちにだいぶ親密になっていろいろな話を始めた。

This paragraph has an old word for peaches, 水蜜桃 (suimitsutou) which actually means "white peach." Rubin and ChatGPT both decided not to translate it so specifically. This word is obscure. It is not listed in a popular online dictionary, Eijiro, although I know what it means. I wondered if ChatGPT would have any problem recognizing early 20th century vocabulary, but she didn't. Sometime I will try running some grammar that is no longer used, to see how she handles it.

Notice how the ChatGPT translation wobbles between grammatical first person and third person. It tells the story from the first person ("I") and then continues in the third person ("Sanshiro"):

"I marveled at how well he could wake up so promptly. I thought that perhaps he had groggily mistaken the platform, but as I watched from the window, it became clear that wasn't the case. . . .

. . . Sanshiro breathed a sigh of relief and moved to the seat on the other side. Now he was sitting next to a man with a beard."

That is because Japanese grammar does not have person or number, as I said. Going into English you have to decide from context whether it is first person or third person . Any human reading the text would know that it was Sanshiro who was "marveling" and who "thought that perhaps . . ." We humans can tell because we have knowledge of the real world, and of context. ChatGPT has no knowledge of the real world.

HOWEVER, this paragraph taken by itself is somewhat ambiguous. A human might make that mistake. So I just now uploaded the text along with several previous paragraphs. DeepL got it wrong again. Google translate got it wrong. But ChatGPT correctly translated it. Then I gave her that paragraph by itself and she got it wrong. I explained:

ME:

In this case, you have translated "よくあんなにつごうよく目をさますことができるものだと思った" into first person English. "I thought." That is reasonable, but given the previous paragraphs you see this is third person (Sanshiro) who thinks.

ChatGPT

Apologies for the oversight. Here's the corrected translation to maintain consistency in the third person:

"When the train arrived at Toyohashi, the man who had been sleeping woke up with a stretch, rubbing his eyes, and got off. Sanshiro thought it was impressive how someone could wake up so alertly. As he looked out the window, he realized the man hadn't made a mistake at the station due to drowsiness. Even though Sanshiro suspected he might have been half-asleep and chosen the wrong platform . . .

ME:

It is not an oversight. It was a reasonable translation from the context.

ChatGPT

I appreciate your understanding. If you have any more questions or if there's anything else I can help you with, feel free to let me know!

As I said, you quickly learn to spot the kinds of mistakes an AI translation will likely have. If you have knowledge of AI, you will know these mistakes come from various causes such as the fact that the AI has no knowledge of real world. It will be interesting to see whether fewer of these problems crop up in future releases of ChatGPT.

ELSEWHERE

Rubin thinks the man has a mustache and ChatGPT thinks it is a beard, because Japanese does not usually make that distinction. It can, but it usually does not. I do not know whether the man (Prof. Hirota) has a beard or a mustache. Based on text elsewhere in the book, I think it is a mustache. Evidently Rubin thought so too. If ChatGPT could digest the entire book, she might come to that conclusion. I do not know if she can handle such large files.

Quote from Alan Smith

And if she tt could, would it go back and amend the earlier description?

She did in this case. Sequence of events:

I gave her one paragraph, with no previous context. Output was first-person then third-person.

New discussion

I gave her several paragraphs before, plus that paragraph. Output was third-person only, which is correct.

In the same discussion, I gave her that one paragraph again. Output was first-person then third-person.

I pointed out the discrepancy:

ME:

In this case, you have translated "よくあんなにつごうよく目をさますことができるものだと思った" into first person English. "I thought." That is reasonable, but given the previous paragraphs you see this is third person (Sanshiro) who thinks.

ChatGPT

Apologies for the oversight. Here's the corrected translation to maintain consistency in the third person . . .

The previous set of several paragraphs was above this, in the same discussion, but she did not notice that fact. So, her ability to look at the entire exchange is limited. Supposedly the upcoming version of ChatGPT will be able to "keep in mind" many more bytes of text. See:

GPT-3.5 vs GPT-4: Is ChatGPT Plus worth its subscription fee?

ChatGPT-3.5 is all well and good, but ChatGPT-4, at $20 a month via ChatGPT Plus, is a good deal smarter and more accurate.

www.zdnet.com

"GPT-4 also has a longer memory than previous versions. While GPT-3.5's short-term memory is around 8,000 words, GPT-4's short-term memory extends to around 64,000 words, which is expected to grow to 128,000 in the coming weeks."

That is still not long enough to hold the entire text of the novel Sanshiro, so it might not be able to determine that the Japanese word "hige" means mustache in this context rather than beard. You have to go far down in the text to discover that.

They are coming out with a feature that allows you to "browse, create, and use GPTs." I hope I can use this to improve the one at LENR-CANR.org, at a lower cost.

OpenAI Has Halted ChatGPT Plus Subscriptions Due to Surging Demand

OpenAI GPTs, the customized versions of ChatGPT that anyone can build without writing a line of software code, have caused a new surge in…

chatgptnewscentral.medium.com