Using AI (ChatGPT) for learning Chinese

[Jan Finster]

On 5/17/2023 at 2:39 PM, Lu said:

And the Summer Palace is 颐和园. Calling it 夏宫 is really weird.

?

User: please translate: "I like going to the summer palace" 

ChatGPT: 我喜欢去颐和园。(Wǒ xǐhuān qù Yíhéyuán.)

 

actually 夏宫 seems to be a common way to call some Russian summer palaces...

 

On 5/17/2023 at 2:39 PM, Lu said:

Using it for language learning, without a more knowledgable person to help you tell correct from incorrect, good from bad and modern from stiff/formal, is not the best idea. That includes vocabulary lists, usage explanations, usage examples... If the sentence-generating machine is right 95% of the time, but you trust it 100% of the time, one in twenty things you learn is wrong and who knows when you'll find out.

 

This assumes you memorise 100% of what you learn. Realistically, while learning you will have to be exposed to the same words in different contexts many times. I do not believe humans are used to having a perfect teacher anyway. Were your parents omniscient and perfect? You probably realised at some point that things you learned as a child from your parents or friends were wrong...and you re-learned.

 

Obviously using ChatGPT is a matter of weighing pros and cons. And yes, I agree, if you are a total beginner and naively trust it, then it can do harm. For me, it has been a godsend. I have been studying Chinese since 2019 but have 99% focussed on input. Finally, I have someone to practice writing/chattting with. I have done so with a teacher a few times, but this is not the same. When you write homework for a teacher it is maybe 20 sentences. In an hour with ChatGPT I can write many more than that. I believe good prompt engineering is important and so is remaining sceptical of any corrections. However, since starting to "chat" with ChatGPT my active vocabulary has grown rapidly. I ask it, for instance, to list 25 attributes of a good teacher. Then I practice making sentences containing those attributes. Next session I might ask it to list 25 expressions related to visa applications. Then I practice making sentences...

 

As a professional translator, I can see how you, Lu, are still sceptical. It is probably not yet at a level suitable for you. However, just as it cannot be use by programmers to write perfect code without review, it should be able to be a great assistant. Many coders approve of its help (while I as a non-coder have not been able to get it to write any code that actually runs).

[roddy]

I’m a professional translator and I fully expect the new generation of tools - NOT the current free ones you can play with - to wipe out a lot of translation work. And not even by doing the translation, exactly. It might be doing the writing directly into the necessary languages. ”Here’s a JSON file of facts about changes to our online banking system. Produce an email in each of the following languages communicating these to users.”

 

And yes, it’ll still need monitored and checked. But I don’t think people have quite got their heads round the fact that computers can now or very soon will be able to, for most practical purposes, understand, process and produce human language, at great speed and a decent to very good degree of accuracy. 
 

It’s easy to say it can’t do a novel or translate an idiom. But the bulk of writing and translating is everything else: internal and external comms, reports, news, product descriptions, price lists. Tedious, repetitive, paying work for a bazillion people. 

[Lu]

On 5/17/2023 at 9:32 PM, roddy said:

I’m a professional translator and I fully expect the new generation of tools - NOT the current free ones you can play with - to wipe out a lot of translation work. (...)
 

It’s easy to say it can’t do a novel or translate an idiom. But the bulk of writing and translating is everything else: internal and external comms, reports, news, product descriptions, price lists. Tedious, repetitive, paying work for a bazillion people.

In that sense it will just continue what has already been happening since Google Translate. A lot things don't need to be translated anymore, because the reader doesn't need an exact translation, they just want to get the general idea. A lot of other text is translated using a mix of translation databases, machine translation and a little human tweaking (the entire Microsoft help pages). Expectations among readers are lowering as well, is my impression. This will continue, but I don't think there will be a steep drop, because it has been going on for a while already.

 

And then there is the question 'which languages'. Translations from English into Western European languages, or vice versa: sure. Translations Korean into Chinese: hahaha no, and this will not change in the forseeable future.

 

But feel free to point & laugh in five or ten years if I'm wrong.

[Jan Finster]

On 5/17/2023 at 9:51 PM, Lu said:

Translations Korean into Chinese: hahaha no, and this will not change in the forseeable future.

Why not? Because the current LLMs are US based and mostly fed English? It will only be a matter of time until Asian countries will catch up. There is a lot of Mandarin and Korean to teach an LLM on. So, I do not see why it should not work.

 

 

On 5/17/2023 at 9:32 PM, roddy said:

But the bulk of writing and translating is everything else: internal and external comms, reports, news, product descriptions, price lists. Tedious, repetitive, paying work for a bazillion people. 

 

Yes, in some way this is great. I do wonder however, how it will affect the current and future generation's choice of career (?) In the past students might have said "I am good with language, I want to become a journalist/teacher/author/etc", now they might feel as if ChatGPT and other LLMs jeopardize their futures. And in some way they do. Being able to write good prose or text is no longer an indispensable  skill ["we do not need a person to do that"]. 

[Lu]

On 5/18/2023 at 9:27 AM, Jan Finster said:

Why not? Because the current LLMs are US based and mostly fed English? It will only be a matter of time until Asian countries will catch up. There is a lot of Mandarin and Korean to teach an LLM on. So, I do not see why it should not work.

Yes, because the current and past LLMs have all been US based and fed on English, and are built by organisations that are monolingual and see English as the only real language (I'm exaggerating a little, but not much). Asian countries could catch up, but they aren't. I'm again looking at Google Translate, which is now over fifteen years old and still bad at Mandarin.

 

I'm sure an interesting socio-linguistic-economic study can be done as to why Asian countries aren't building their own language and translation systems. India could do with one, with its myriad languages. Mandarin is so big it should have one. No doubt money is part of the reason, and the people who would be best at this work leaving for Silicon Valley to work there, in English. And perhaps a smattering of the notion that Chinese is so rich and complex that no machine could fully process it. What Mandarin does have is a good transcription program, working in Wechat, where it handily transcribes voice messages.

 

On 5/18/2023 at 9:27 AM, Jan Finster said:

I do wonder however, how it will affect the current and future generation's choice of career (?) In the past students might have said "I am good with language, I want to become a journalist/teacher/author/etc", now they might feel as if ChatGPT and other LLMs jeopardize their futures. And in some way they do. Being able to write good prose or text is no longer an indispensable  skill ["we do not need a person to do that"]. 

I worry about this too. To get good at language, you have to write a lot, polish your writing, write more. I think there will always be some demand for excellent writing skills, but how are people supposed to get excellent at writing if the reasonably-okay writing they could practice with is taken over by LLMs?

[Apoc]

@Lu Forget Google Translate, you should really give ChatGPT a try. In my experience Korean-Mandarin translation is already 95% perfect, with the remaining 5% being mostly easy-to-fix stylistic mistakes. Mileage will vary of course depending on the field, my experience is from translating economic jargon.

[Johnny-5]

On 5/15/2023 at 5:50 AM, TheBigZaboon said:

Many, many years ago I took a graduate course in Chinese history, with his son as a fellow student.

heh, that's pretty cool 

[Johnny-5]

On 5/15/2023 at 10:34 AM, Moshen said:

 

This happened to me, too.  It was so strange because the authors were known to me and the article titles seemed very plausible for them but the articles actually did not exist.

Yeah, it is disturbingly good at making things up that are completely untrue. Some may even call it a "lie" 

 

There is a dystopian future where all of the things that people believe are extremely convincing lies generated by a computer.

[roddy]

I suspect translation between the 'harder'/smaller languages will see big boons once the LLMs are produced, as the existing bridge translation through English will no longer be needed. And there is no way the Chinese government is going to pass up better monitoring of speech. They might get there slowly, but they'll get there.

[Apoc]

I like that you put 'harder' in quotes. I believe that LLMs will show that all languages are pretty much equally difficult, the only thing that matters to AI is the quality and quantity of training data.

[TheBigZaboon]

@Johnny-5 Another very barely on-topic interruption, and only justified because you dropped his name, but you have an English-style reference to Rampo Edogawa. But using his name like that wastes his efforts to honor his idol and muse as a mystery writer. He chose his pen name, Edogawa Rampo (as it's written in Japanese) to honor Edgar Allen Poe...

 

Sorry, another uninvited interruption...

 

TBZ 

[Lu]

On 5/18/2023 at 1:41 PM, Apoc said:

I like that you put 'harder' in quotes. I believe that LLMs will show that all languages are pretty much equally difficult, the only thing that matters to AI is the quality and quantity of training data.

I'm happy to accept that all languages are equally difficult, but they are not all equally difficult to translate from and into. The Western European languages are easier to translate into one another than into or from Chinese, for one. English <-> Chinese is a bit easier than Dutch <-> Chinese, especially when it comes to realia, because so many Chinese things already have a fairly established English name but no Dutch name, and so many cultural English things are already known and named in Chinese. I guess you could file that under 'quality and quantity of training data', but then it is just harder because of the difference in training data.

 

On 5/18/2023 at 11:16 AM, Apoc said:

Forget Google Translate, you should really give ChatGPT a try. In my experience Korean-Mandarin translation is already 95% perfect, with the remaining 5% being mostly easy-to-fix stylistic mistakes. Mileage will vary of course depending on the field, my experience is from translating economic jargon.

If you say so! I can't judge its qualities in Korean, and economic texts does sound like a field that has language that is high on jargon and low on grammatical intricacies that make sentences difficult.

[Apoc]

You don't need to take my word for it, it's free to try. Just wanted to let you know my experience.

[Apoc]

On 5/18/2023 at 3:53 PM, Lu said:

English <-> Chinese is a bit easier than Dutch <-> Chinese, especially when it comes to realia, because so many Chinese things already have a fairly established English name but no Dutch name, and so many cultural English things are already known and named in Chinese. I guess you could file that under 'quality and quantity of training data', but then it is just harder because of the difference in training data.

This is a good point. Intuitively I would think that a human translator will have an edge over AI if there's a big "gap" between the source and target language, since more creativity will be needed to make the translation sound natural. In practice I haven't noticed ChatGPT having much trouble filling those gaps though. Besides Korean-Mandarin, I have translated a lot to English from Finnish (my native language, very far apart from English). It's just incredibly good.

[Lu]

On 5/18/2023 at 5:14 PM, Apoc said:

Intuitively I would think that a human translator will have an edge over AI if there's a big "gap" between the source and target language, since more creativity will be needed to make the translation sound natural.

I have the impression that ChatGPT is quite good at producing natural-sounding English. The challenge lies in producing a translation that sounds natural and is at the same time as close as possible to the original. (I mainly translate literature, so that is on my mind. I realise that the situation will be different for many commercial texts.)

[realmayo]

I'm blown away by its translation abilities. Rather than ploughing through a technical text in Chinese it's way easier just to skim the translated version.

 

e.g.

 

Spoiler

1. Functional Characteristics of "Shi...de" Sentences (2)

"Shi...de" sentences (2) refer to a part of verb predicate sentences and adjective predicate sentences that carry the "shi...de" structure. Both "shi" and "de" indicate mood. These types of sentences are often used to express the speaker's evaluation, description, or depiction of the subject. The entire sentence often carries a positive tone aimed at explaining the situation, elaborating on a point, or trying to persuade or convince the listener.

2. Structural Characteristics of "Shi...de" Sentences (2)

In "Shi...de" sentences (2), we can consider "shi" as a modal adverb, usually used before the predicate, and "de" as a modal particle used at the end of the sentence. The predicate in the middle of "Shi...de" can be a verb or an adjective. Common verb predicates are often in the form of "nengyuan verb + verb" or "verb + possible complement." Fixed phrases that function similarly to adjectives can also be used in "Shi...de" sentences (2).

Verbs expressing mental activities and emotions can also be used in "Shi...de" sentences (2). These verbs can be modified by degree adverbs, and the object of the verb often appears at the beginning of the sentence as the topic.

1. Regarding the Use of "Shi...de" Sentences (2)

Since "Shi...de" sentences (2) express a positive tone in their expression, they generally do not have a negative form like "bu shi...de" (not being). However, negative expressions can be used within the "Shi...de" structure. In such cases, the overall sentence still carries a positive tone.

3. Omission of "Shi" and "De"

In "Shi...de" sentences (2), except for some sentences with "zhe" or "na" as the subject and sentences expressing double negation, both "shi" and "de" can be omitted simultaneously, or only "shi" can be omitted. The meaning of the sentence remains the same, but without "shi...de," the sentence no longer belongs to "Shi...de" sentences (2) and becomes a general verb or adjective predicate sentence. The tone also naturally changes. When using "shi...de," the tone is affirmative, polite, and gentle, conveying a sense of reasoning to make people believe. Without "shi...de," the tone becomes stronger and sometimes appears concise and straightforward.

[roddy]

Or skip the skimming and have it summarise. Or if you know what you’re looking for, ask it to do the skimming. 
 

Interesting news on coverage of smaller languages. 
https://about.fb.com/news/2023/05/ai-massively-multilingual-speech-technology/

[realmayo]

I'm starting to wonder if there won't be any long-lasting AI language learning programmes, for the simple reason that most people will stop learning languages.

 

If you go to a foreign country you will surely be able to just wear a regular bluetooth earphone with microphone: when you speak your phone will output your words into the local language, and send any replies into your ear, translated into your own language.

 

I assume this technology already exists.

 

 

[markhavemann]

On 5/28/2023 at 1:14 AM, realmayo said:

the simple reason that most people will stop learning languages.

All kinds of different forms of transportation exist but people still choose to ride bicycles and walk. People still read paper books despite Kindles existing. Computers and phones didn't make pens and paper disappear. Painting and drawing still happens despite having electronic drawing tablets.

 

Usually new technologies don't kill the original activity, it just gives people a choice. I think most people who learn languages as adults mostly do it for other reasons than absolute necessity. 

 

Edit: I just saw that this has forked into a thread of it's own. Can't delete a post though.

[Lu]

On 5/27/2023 at 7:14 PM, realmayo said:

If you go to a foreign country you will surely be able to just wear a regular bluetooth earphone with microphone: when you speak your phone will output your words into the local language, and send any replies into your ear, translated into your own language.

 

I assume this technology already exists.

I've seen some promotion for such a thing some years ago. I was skeptical it could work. I haven't heard of it since. If it existed and worked well enough, I think we'd know.