Mandarin with an English accent
« previous post | next post »
Something very funny happened to me earlier today, funny enough that I would like to share it with all Language Log readers who may be desirous of something more than a cup of coffee to perk them up on a gray, midweek morning.
I entered the following Mandarin expression into Google Translate and wanted to hear it pronounced by the machine: 衷心感謝 ("heartfelt thanks"). So I clicked on the speaker button, but, by mistake, I had it set to English rather than to Chinese. What I heard was Mandarin with an English accent!
When set to Chinese, the machine pronounces 衷心感謝 properly and precisely: zhōngxīn gǎnxiè. When set erroneously to English, it sounds like an American reading out romanized Mandarin, with the "correct" suprasegmental intonation and all, but, of course, paying absolutely no attention to lexical tones. Amazingly, it's still understandable, which replicates the experiments my wife used to make by going up to strangers on American streets and asking them to read pinyin Mandarin to native speakers. She was always triumphant when the native speakers could understand most of what the English speakers were reading.
I had the machine read 衷心感謝 in French, Spanish, Italian, German, and other languages, and they all had their own special "flavor".
Then I thought I'd give it a try on a longer passage, so I entered the opening sentences of Zhu Ziqing's famous 1925 essay, "Bèiyǐng 背影" (conventionally rendered as "Retreating Figure", more literally, "a view of somebody's back; a figure viewed from behind; silhouette from the back" [I don't think that anyone else has used this translation for the Zhu Ziqing story before]):
Wǒ yǔ fùqīn bù xiāng jiàn yǐ yǒu èr nián yú le, wǒ zuì bùnéng wàngjì de shì tā de bèiyǐng. Nà nián dōngtiān, zǔmǔ sǐle, fùqīn de chāishǐ yě jiāoxiè le, zhèng shì huòbùdānxíng de rìzi, wǒ cóng Běijīng dào Xúzhōu, dǎsuàn gēnzhe fùqīn bēnsàng huí jiā. Dào Xúzhōu jiàn zhao fùqīn, kànjiàn mǎn yuàn lángjí de dōngxī, yòu xiǎngqǐ zǔmǔ, bùjīn sùsùde liúxià yǎnlèi.
我與父親不相見已有二年餘了，我最不能忘記的是他的背影。那年冬天，祖母死了，父親的差使也交卸了，正是禍不單行的日子，我從北京到徐州，打算跟著父親奔喪回家。到徐州見著父親，看見滿院狼籍的東西，又想起祖母，不禁簌簌地流下眼淚。
Translations:
DeepL
I have not seen my father for more than two years, but what I cannot forget is his back. In the winter of that year, my grandmother died and my father's job was handed over. When I saw my father in Xuzhou, I saw the wreckage in the courtyard and thought of my grandmother, so I couldn't help but shed tears.
Google Translate:
It has been more than two years since I saw my father, and the last thing I can forget is his back. That winter, my grandmother died, and my father's errand was relinquished. It was a day when misfortunes never come singly. I went from Beijing to Xuzhou, intending to follow my father home for the funeral. When I went to Xuzhou and met my father, he saw at least one messy house, and thought of my grandmother, I could not help but shed tears in Susu.
Baidu Fanyi:
I haven't seen my father for over two years, and the last thing I can't forget is his back. That winter, my grandmother passed away, and my father's errand was also handed over. It was a day when disaster never came singly. I went from Beijing to Xuzhou and planned to follow my father's funeral home. When I met my father in Xuzhou, saw the messy things in the courtyard, and thought of my grandmother, I couldn't help but shed tears.
Microsoft Bing:
It's been more than two years since I saw my father, and the last thing I can forget is his back. That winter, my grandmother died, and my father's messenger was handed over, and I traveled from Beijing to Xuzhou, intending to follow my father home from his mourning. When I went to Xuzhou to see my father, I saw the wolf things in the courtyard, and I thought of my grandmother, and I couldn't help but shed tears.
VHM: All these machine translations are defective in one or more ways (most conspicuously the second sentence of the text, which I would translate as "The hardest thing for me to forget about him is his silhouette from the back". So I also quote here the following version from David E. Pollard, "The View from the Rear", in David E. Pollard, trans. and ed., The Chinese Essay (New York: Columbia University Press, 2000), pp. 21-219 of 217-221:
It has been two years and more since I saw my father. My most vivid memory of him is a view of him from the rear.
That winter my grandmother had died, and my father's job had come to an end; our troubles truly did not come singly then. I left Peking for Xuzhou, to accompany my father home for the funeral. When I saw the household things strewn about the yard, and thought too of my grandmother, I wept copiously.
Zhu's "Silhouette from the Back" has been translated into English by Lau & Goldblatt (1995) and Zhang Peiji (1999), and the passage in question is also available scattered about elsewhere in various books and articles.
Selected readings
- "Jichang Lulu" (9/26/22) — be sure to read the opening paragraphs
- "Characterless Sinitic" (9/1/21)
- "Mandarin with a German accent" (5/8/19)
[Thanks to Jing Hu]
Paul Frank said,
April 6, 2023 @ 6:06 am
And here's Chat GPT, with the prompt, render this accurately in fluent English. Don't translate literally:
"It has been more than two years since I last saw my father, and what I cannot forget is the image of his back. That winter, my grandmother passed away, and my father had also lost his job. It was a time of misfortune upon misfortune. I traveled from Beijing to Xuzhou, intending to follow my father back home to mourn our loss. When I arrived in Xuzhou and saw my father, the sight of the messy yard and the thought of my late grandmother brought tears streaming down my face."
Prompt 2: Try again, Chat GPT:
"I haven't seen my father for over two years, and what I can't forget most is his silhouette. That winter, my grandmother passed away, and my father lost his job – it was a time of one misfortune after another. I traveled from Beijing to Xuzhou, intending to accompany my father back home to mourn. Upon arriving in Xuzhou and seeing my father, I was confronted with the disarray of our family belongings. This, combined with the memory of my grandmother, brought tears streaming down my face."
Prompt 3: One last time:
"I haven't seen my father for more than two years, and what I can't forget the most is his silhouette. That winter, my grandmother passed away and my father lost his job. It was a time when misfortunes seemed to come one after another. I traveled from Beijing to Xuzhou, intending to join my father in mourning and return home. When I arrived in Xuzhou and saw my father, the sight of the messy yard, combined with the memory of my grandmother, brought tears to my eyes."
Paul Frank said,
April 6, 2023 @ 6:07 am
And here's Chat GPT, with the prompt, render this accurately in fluent English. Don't translate literally:
"It has been more than two years since I last saw my father, and what I cannot forget is the image of his back. That winter, my grandmother passed away, and my father had also lost his job. It was a time of misfortune upon misfortune. I traveled from Beijing to Xuzhou, intending to follow my father back home to mourn our loss. When I arrived in Xuzhou and saw my father, the sight of the messy yard and the thought of my late grandmother brought tears streaming down my face."
Prompt 2: Try again, Chat GPT:
"I haven't seen my father for over two years, and what I can't forget most is his silhouette. That winter, my grandmother passed away, and my father lost his job – it was a time of one misfortune after another. I traveled from Beijing to Xuzhou, intending to accompany my father back home to mourn. Upon arriving in Xuzhou and seeing my father, I was confronted with the disarray of our family belongings. This, combined with the memory of my grandmother, brought tears streaming down my face."
Victor Mair said,
April 6, 2023 @ 6:32 am
Astonishing!
Almost superhuman!
Not only can the chatbots answer questions, hold conversations, and tell tales, now we find that apparently they can also make excellent translations.
,
Paul Frank, you use the prompts "render" and "translate", so it sounds like the bot is working from Chinese to English, but is that actually the case? Or rather, is it amalgamating and "correcting" the faulty AI translations I provided.(perhaps relying heavily on the human one from Pollard) as well?
No matter how the chatbot accomplished the feat, the stellar quality of the multiple versions produced by it is mind-boggling.
Paul Frank said,
April 6, 2023 @ 6:50 am
I entered the Chinese. And I can ask the Chatbot to give me 10 or 20 different translations, and also to tweak individual translations this way or that. It will make mistakes, sometimes egregious ones, but it also gets a lot of things astoundingly right. The key to using the tool is the formulation of prompts. From last week's Economist: "Amid the hype, it is easy to forget a minor miracle. Chatgpt has aced a problem that long served as a far-off dream for engineers: generating human-like language. Unlike earlier versions of the system, it can go on doing so for paragraphs on end without descending into incoherence. And this achievement’s dimensions are even greater than they seem at first glance. Chatgpt is not only able to generate remarkably realistic English. It is also able to instantly blurt out text in more than 50 languages—the precise number is apparently unknown to the system itself. Asked (in Spanish) how many languages it can speak, Chatgpt replies, vaguely, “more than 50”, explaining that its ability to produce text will depend on how much training data is available for any given language. Then, asked a question in an unannounced switch to Portuguese, it offers up a sketch of your columnist’s biography in that language. Most of it was correct, but it had him studying the wrong subject at the wrong university. The language itself was impeccable.
Portuguese is one of the world’s biggest languages. Trying out a smaller language, your columnist probed Chatgpt in Danish, spoken by only about 5.5m people. Danes do much of their online writing in English, so the training data for Danish must be orders of magnitude scarcer than what is available for English, Spanish or Portuguese. Chatgpt’s answers were factually askew but expressed in almost perfect Danish. (A tiny gender-agreement error was the only mistake caught in any of the languages tested.)
Indeed, Chatgpt is too modest about its own abilities. On request, it furnishes a list of 51 languages it can work in, including Esperanto, Kannada and Zulu. It declines to say that it can “speak” these languages, but rather “generates text” in them. This is too humble an answer. Addressed in Catalan—a language not on the list—it replies in that language with a cheerful “Yes, I do speak Catalan—what can I help you with?” A few follow-up questions do not trip it up in the slightest, including a query about whether it is merely translating answers first generated in another language into Catalan. This, Chatgpt denies: “I don’t translate from any other language; I look in my database for the best words and phrases to answer your questions.”
Who knows if this is true? Chatgpt not only makes things up, but incorrectly answers questions about the very conversation it is having. (It has no “memory”, but rather feeds the last few thousand words of each conversation back into itself as a new prompt. If you have been speaking English for a while it will “forget” that you asked a question in Danish earlier and say that the question was asked in English.) Chatgpt is untrustworthy not just about the world, but even about itself.
This should not overshadow the achievement of a model that can effortlessly mimic so many languages, including those with limited training data. Speakers of smaller languages have worried for years about language technologies passing them by. Their justifiable concern had two causes: the lesser incentive for companies to develop products in Icelandic or Maltese, and the relative lack of data to train them."
Victor Mair said,
April 6, 2023 @ 7:28 am
@Paul Frank
Your last comment, for which thanks, ends with a double quotation mark. Where does that quotation begin?
Paul Frank said,
April 6, 2023 @ 8:33 am
It starts with "Amid hype". ChatGPT, as everyone knows by now, makes all sorts of mistakes and is prone to spewing unadulterated bullshit (called "hallucinations" in the jargon), but is nevertheless remarkably proficient at producing fluent language. The day will come when we will be able to upload termbases for specific terminology as well as specific style guides (or pick them from a dropdown menu) and post-edit texts translated by these tools to produce excellent translations much more quickly than is possible today. But you will still need humans who know the source language and have native-level skills in the target language to check the accuracy of the translations. Or perhaps not: I can imagine a future when hundreds of tools working simultaneously check the output produced by one machine and flag and correct errors identified by a majority of them.
Mark S. said,
April 6, 2023 @ 8:52 am
The readings produced are not entirely unlike those produced intentionally by humans in the early days of Taipei's subway system. Back then, station names were labeled in Chinese characters and bastardized Wade-Giles (e.g., what would be "Guting" in Hanyu Pinyin was identified as "Kuting"; and what should be "Da'an" in Hanyu Pinyin was called "Taan").
What was worse than the spellings, though, was the announcer's "English" pronunciations of the station names. Each was read out in an exaggerated and incorrect way as if pronounced by the world's most clueless and linguistically inept tourist. What should have been pronounced "Dà'ān" instead was given as something more like "Tāh-ànn," with the second "a" as in "apple" (Mandarin doesn't have that sound). The city's mass rapid transit system delivered an auditory assault upon riders, and it took years of complaints by pained foreign residents before the city finally fixed that.
And although the city finally adopted Hanyu Pinyin and adopted more or less proper Mandarin pronunciations in the English announcements, nearly twenty years later Taipei and New Taipei still haven't fixed the errors in Taipei's MRT system signage caused by missing apostrophes. And it's not like I haven't told them … repeatedly.
Mike Grubb said,
April 6, 2023 @ 9:06 am
ChatGPT is no doubt an impressive accomplishment, but I would interrogate the description that it achieves output "effortlessly." It may be fast and easy at the front end for users, but a lot of contractor labor, often by workers in developing economies, goes into filtering the training data so that its outputs are less likely to be problematic. "Seemingly effortlessly" might be more accurate.
Paul Frank said,
April 6, 2023 @ 9:53 am
Mike, that's true of just about every product produced in the neoliberal ecumene that now covers our entire planet. Nonetheless, it is astounding. As a native speaker of Chilean Spanish and a resident of Bern, I'm floored by this output:
Me: Render this in natural sounding Chilean Spanish, with Chilean colloquialisms, and in German with a Swiss flavor:
Chilean Spanish:
Sin duda, ChatGPT es una maravilla, pero yo le pondría ojo a la descripción de que logra resultados "sin esfuerzo". Puede que sea rápido y fácil pa' los usuarios en la parte de adelante, pero hay harta pega de contratistas, muchos de ellos en países en vías de desarrollo, que se dedican a filtrar los datos pa' entrenarlo y así sus respuestas sean menos cachosas. "Aparentemente sin esfuerzo" sería más preciso, cachai?
Swiss German:
ChatGPT isch sicher es beeindruckends Eruugnis, aber ich würd d'Beschriibig, dass es die Usgab "ohni Müeh" erreicht, hinterfrage. Es cha schnäll und eifach für d'Nutzer am Afang si, aber es het viel Arbet vo Uftragsarbeiter inne, oft us Entwickligsländer, wo i d'Filterig vo de Trainigsdate investiert wird, damit d'Usgab weniger problematisch sind. "Schynbar ohni Müeh" wär villicht zutreffender.
cameron said,
April 6, 2023 @ 10:32 am
It looks like the quotation mark at the end of Paul Frank's comment above matches the one that follows "From last week's Economist:"
Those are the only instances of " in the comment. the other quotation marks are the curly kind, “ and ”
Jim Unger said,
April 6, 2023 @ 1:42 pm
I think Google Translate assumes that the selected (or "detected") language parameter defines the rules of orthography to use in producing a syllabic parsing of the input text (pinyin in the case of Chinese). The synthesized voice then uses preprogrammed rules of orthography/phonology correspondence to crank out a stream of syllables, which it smooths out before final output in accord with the typical pitch contours of the specified language. If the result of choosing English to pronounce Chinese sounds impressive, I think it is mostly because of the limited range of orthographic variation in pinyin syllables. Try inputting a real French sentence and getting the program to say it "in English": the result is much worse than even a beginning American student would produce. Surely this is because both English and French orthographies deviate a lot from letter-phoneme biuniqueness.
KWillets said,
April 6, 2023 @ 3:02 pm
It appears to be reading the romanized version of the expression. In Korean at least it's audibly using the Revised Romanization without the specialized vowels.
너구리
(However when I put the pinyinized Chinese back in, I get a different pronunciation from the original characters.)
Jonathan Smith said,
April 6, 2023 @ 3:20 pm
We discussed ChatGPT's translation skills wrt Chinese extensively in previous threads… or rather I did :D
"最不能忘記的是X"
hmm across the board fail here… try "the one thing I can't forget" "the hardest thing to forget" or use "unforgettable", etc., anything normal works
you can tell ChatGPT its translation sucks and why; it will agree and revise. Can it disagree? I wonder
Terry K. said,
April 6, 2023 @ 4:06 pm
After reading Jim Unger's comment, I tried it with a trio of simple French phrases, with it set to English. And it did okay, to my ear (as someone who knows those phrases but doesn't speak French) on most of the words. But if the word was (as written) also an English word, it mangled it. That is, it mangles homographs. Mandarin, written in characters, has the advantage of not having homographs with English.
Paul Frank said,
April 6, 2023 @ 6:31 pm
Jonathan: I don't use ChatGPT to translate anything. But I do use it to produce text for me in various languages I know to various degrees based on text I input in English. I then correct mistakes or delete phrases I don't want. Used in this way, it produces much better German or French or, yes, Chinese than DeepL. That said, DeepL does have the advantage of drop-down menus with synonyms and a dynamic ability to alter the rest of sentences when I start rewriting the beginning of the initial output sentence. CAT tools already incorporate DeepL alongside terminology databases and translation memories that can be managed professionally. Pretty soon, they will also incorporate ChatGPT. And as I mentioned earlier: you can ask ChatGPT to give you 10 or 20 different versions of a paragraph and tweak the output by prompting it within a single conversation. The prompt: "translate this" is almost always sure to produce unsatisfactory results.
Jonathan Smith said,
April 6, 2023 @ 7:12 pm
@Paul Frank
yes exactly what I was pointing out 2/28 e.g.
[…]
ChatGPT: "Yes, you are correct that the translation of '人类仍然是掌握和创造语言的主体' as 'humans are still the main body of mastering and creating language' is not the most accurate translation. A more accurate translation might be: 'Human beings remain the primary agents in mastering and creating language.'
me: "very nice except that 'mastering' for 掌握 seems odd here…"
chatgpt: "You are correct, the word 'mastering' may not be the best choice for translating 掌握 in this context. A more appropriate translation might be 'understanding' or 'commanding'
[…]
I recently tried DeepL on some technicalish English > Chinese and found it not good enough even to be slightly helpful, whereas the GPT tools are obviously great translation ASSISTANTS as you note
Paul Frank said,
April 6, 2023 @ 9:33 pm
@Jonathan Smith. Another important thing to keep in mind, which many people don't, is that to use Chat GPT (or a machine translation tool such as DeepL) effectively in order to translate texts, you need to not only post-edit the machine's output but also _pre-edit_ the source text. Changing one word or two or three in the source text (or prompt in the case of Chat GPT) is often all you need to do to help the machine help you produce impeccably written texts in another language that mean exactly (or very close to exactly) what you want them to mean. The advantage of Chat GPT is that you can do this by talking with the machine, i.e. through prompts that prod and nudge it in the direction you want it to go.
liuyao said,
April 15, 2023 @ 6:29 am
Did you not ask Google to read the passage with English accent?