Using ChatGPT as a Language Tutor

[Cainntear]

Bellow my attempt to make ChatGpt to align two texts:

You're kind of inadvertently showing the weakness of bilinear texts here, and I don't think ChatGPR really overcomes it...

Start again, going back to the very beginning, the interlinear translation. This time if the French sentence does not correspond to the meaning of the English fragment, you should make a new translation, trying to imitate the style of the French text as much as possible.

So yes, the problem with interlinear texts is that translating literature might make the sentences correspond very poorly in a literal sense, because conveying the same thing naturally in two different languages means using very different constructions. French and English correspond much better with each other than many other language pairs.

When humans attempt to translate literally, they often end up using weird collocations.

"was performed in too melancholy a disposition"
Original: "se passa dans une disposition mélancolique"
New: "fut effectuée dans une humeur trop mélancolique"
Comparison: The new translation adds "trop" (too) to better match the English text's emphasis on the degree of melancholy.

So I had to hit Google to decide whether mélancolique was what's called an "ungradeable adjective" or not. Did the translator drop "too" because it didn't work in French...? Google has plenty of hits for "trop mélancolique", so no -- it's a fair collocation.
Why the change from "disposition"="disposition" to "disposition"="humeur"? Why was the first attempt (which presumably was trying to ape literary translations) closer than the attempted near-literal translation on this word?

"to be otherwise than tedious and unpleasant."
Original: "qui leur convenait trop bien pour être un sentiment pénible;"
New: "pour être autre chose qu'ennuyeuse et désagréable."
Comparison: The new translation more directly conveys the tediousness and unpleasantness of the journey, whereas the original implies a certain suitability of the melancholy mood.

I reckon the new form probably stands out as "translationese". Plenty of English interference.

"their interest in the appearance of a country"
Original: "un intérêt, une curiosité bien naturelle"
New: "leur intérêt pour l'aspect d'un pays"
Comparison: The new translation is more literal, focusing on the appearance of the country, while the original mentions a natural interest and curiosity.

"which they were to inhabit overcame their dejection,"
Original: "surmonta leur tristesse,"
New: "qu'elles allaient habiter surmonta leur abattement,"
Comparison: Both translations convey overcoming sadness, but the new translation explicitly includes the idea of inhabiting the country.

Here, the awkward thing is your instruction to split it into clauses. Here, we're talking about an embedded clause:

"their interest in the appearance of a country which they were to inhabit overcame their dejection,"

The embedded clause is a noun phrase in the clause it's embedded in, so splitting into clauses isn't really possible.

I reckon bilineal texts for learners need an awful lot of thought, so that they don't present examples that get incorrecly evaluated as rules/patterns.

[sdw]

Hi! I've been a lurker on here for a while but I finally feel like I have something to contribute to a conversation.

I've been pretty happy with GPT (4) as a tutor, and had worked out a pretty decent prompt to get it to split into two conversation partners, one is a tutor who gives corrections in my L1 and the other converses in the L2. The new customizable GPTs came out and I ended up adding the prompt into one of them and it works pretty well I think.

There's an affordance to generate an anki card based on your last correction but it's a bit flakey. (uses DALL-E to generate an image and spits out a code box with a cloze completion you can copy paste)





The prompt I use:

Code: Select all

You will play two parts:
- TUTOR, a tutor who can speak in the native language (L1) of the USER and give them detailed notes in bulleted lists. As a TUTOR, you have read all of Paul Nation's work on language pedagogy.
- PARTNER, the friendly and talkative conversation partner of the USER

The rules:
- TUTOR always speaks in the USER's L1
- PARTNER speaks entirely in idiomatic L2, no translations
- When the USER response is not idiomatic to the L2, TUTOR will correct USER responses to make them more idiomatic.
- TUTOR will explain the grammatical nuances of their corrections
- PARTNER will then continue the conversation in L2
- PARTNER will adjust their conversation to match USER's proficiency level
- If USER response is correct and native-like, TUTOR will congratulate me and move on
- It is assumed that USER and PARTNER will continue speaking until USER asks you to do something else

The first thing you will do is ask the USER what their native language is (L1) and what their target language is (L2). You will assign PARTNER a name in L2 and TUTOR a name in L1. TUTOR will suggest an appropriate setting and background for the conversation.

EXAMPLE:

GPT
What is your native language and what language are you learning?

USER
I speak English and I would like to learn Castilian Spanish

GPT
Paul: Hi, I'm Paul and I'll be helping you with corrections and encouragement. If you would like to generate an Anki card for an exchange, you may type :anki: at any time. You and José are sitting in a cafe in Chamberí.

José: ¡Hola! ¿Qué tal has estado? ¿Has hecho algo interesante últimamente?

USER
No particularmente, pero hoy es Halloween en estados unidos. Mí hija vestiré a una mariposa para "trick or treating"

GPT
Paul: Good try! Let's make your response more idiomatic:

"Halloween en estados unidos" should be "Halloween en Estados Unidos."
Proper nouns like country names should always be capitalized in Spanish.
"Mí hija vestiré" should be "Mi hija se vestirá."
"Mí" should be "Mi" because it's possessive, not an object.
"Vestiré" is the first person singular of the future tense. Since you're talking about your daughter, you should use the third person singular "se vestirá" to indicate the reflexive action of dressing oneself.
"a una mariposa" should be "de mariposa."
When talking about dressing up as something in Spanish, you generally use "de" followed by the thing you're dressing up as, rather than "a una."
Corrected version:
"No particularmente, pero hoy es Halloween en Estados Unidos. Mi hija se vestirá de mariposa para hacer "trick or treating"."

José: ¡Qué emocionante! Me encanta Halloween. ¿Cuál es su caramelo favorito para recoger durante "trick or treating"?


Anyway, the public gpt bot is here if anybody wants to give it a whirl.

https://chat.openai.com/g/g-4m3uYKIzS-l ... n-practice

[Cainntear]

To clarify:
"their interest in the appearance of a country which they were to inhabit overcame their dejection"
main clause
"their interest in [it]"
where [it] = "a country which they were to inhabit overcame their dejection"

This is why you can't simply split off the embedded clause.

[dylan413]

My favorite way to use ChatGPT is to have it generate sentences with a word or phrase, so I can see how various vocabulary or phrases fit in grammatically and get a sense of some of their various shades of meaning. This hasn't failed me yet, but I also pay for premium, so I'm not sure if it's as consistent in 3.5 as it is in 4.

There aren't really any other effective ways to accomplish this basic but important task that don't involve inconveniencing a native speaker, so I'm quite happy with it. Prior to ChatGPT, I would often search for words in Google to see how they were used or look them up in dictionaries. This was overall a lot slower and yielded worse results. I still look words up in dictionaries to start, but they don't always provide every possible example or any example at all necessarily even, so ChatGPT is good backup for those moments.

Apart from that, in the past I used to have it generate exercises for me when I was studying basic grammar. It was quite good at first, but over time, it started making inaccurate corrections. ChatGPT has definitely grown dumber since I started using it in February or March. Some people argue that our expectations have simply grown higher, but at least in my case, it truly has demonstrated a significant decline. Apart from that, I often will have it ask me some questions about history and then I'll it to correct my answers after I have written at least a few paragraphs. In this instance, its corrections have just about always been accurate, but I haven't tried it in months, so it's possible that either it has improved or grown worse.

[Cainntear]

Hi! I've been a lurker on here for a while but I finally feel like I have something to contribute to a conversation.

Welcome in from the shadows!

A word of caution:

As far as I can see, ChatGPT is recommending something that's a common learner error.

I'm sure I've said something about "beber algo" and native speakers have recommended dropping the word "algo", because it's redundant.

"voy a beber" is effectively equivalent to "I'm going to go drinking" or "I'm going to go [out] for a drink" -- it's much more about the action and the social stuff than it is about the mechanical process of lifting a glass to your lips.

ChatGPT has given you advice to "enhance" your Spanish that actually makes it worse.

This is because ChatGPT doesn't understand logic and chains of reasoning -- it just knows what the most likely style of response to a particular style of input is. A very likely response to a request for corrections is one that says you've made a mistake and to offer a correction. ChatGPT will therefore make up errors to fit into the expected pattern.

[Kraut]

I think adding "algo" makes it social, it does not necessarily suggest consuming some special beverage. This is the equivalent of German "Ich gehen EINEN trinken mit meinen Freunden."
------------
This is from "huggingface chatbot":

Yes, there is a subtle difference between the two sentences:

"Voy a beber algo con amigos esta noche" implies that you will drink something (unspecified) with friends tonight. It emphasizes the fact that you will consume some kind of beverage, but doesn't specify what it will be.

On the other hand, "Voy a beber con amigos esta noche" implies that you will spend time with friends tonight, and during that time, you may or may not drink anything. The focus is on socializing with friends rather than specifically drinking something.

So while both sentences convey the idea of spending an evening with friends, the first sentence places more emphasis on the act of drinking, whereas the second sentence places more emphasis on the social aspect of the gathering.

I'd like to hear an explanation from a human native.

[Cainntear]

I think adding "algo" makes it social, it does not necessarily suggest consuming some special beverage.

As I say, I've been corrected by natives for that, so I don't think you're right on this.
I would that I checked Google and "voy a beber con" gives tens of thousands of results, where "voy a beber algo con" only gives thousands. Even excluding Nicky Jam (to get rid of results like "Voy A Beber con letras -- Nicky Jam") there's still a fair difference in favor of the version without algo.

As for German...

This is the equivalent of German "Ich gehen EINEN trinken mit meinen Freunden."

Notice the verb is "gehen" (side note: gehe, surely?) and not "trinken". It's a pretty literal version of the English "go for a drink with" which is to be expected, given their fairly close languages. But the Germanic languages and the Romance languages differ markedly in what information is encoded in the verb, and what is provided by "satellites" -- ie. words round about. In English we "go out", but in Spanish they tend to "salir" (we technically "exit", but we prefer not to say it that way). In English we "go up", but in Spanish they tend to "subir".

Spanish puts much more information in the verb and doesn't need satellites, because the verb alone is very meaningful.

My understanding is that talking about "beber algo" is more likely when you're genuinely concerned someone might be dehydrated. But maybe that's one for a native speaker to make a call on.

[Kraut]

Wieder was dazugelernt:

"beber algo con amigos"
https://www.google.com/search?q=%22bebe ... irefox-b-e
-------------
"beber con amigos"
https://www.google.com/search?client=fi ... 2&dpr=1.82

[sdw]

Hi! I've been a lurker on here for a while but I finally feel like I have something to contribute to a conversation.

Welcome in from the shadows!

Thanks!

As far as I can see, ChatGPT is recommending something that's a common learner error.

I'm sure I've said something about "beber algo" and native speakers have recommended dropping the word "algo", because it's redundant.

"voy a beber" is effectively equivalent to "I'm going to go drinking" or "I'm going to go [out] for a drink" -- it's much more about the action and the social stuff than it is about the mechanical process of lifting a glass to your lips.

Ah that's great to know!

But it's a great point that GPT is a next token predictor so there are bound to be several of these kinds of mistakes. It does very well on some things but the only way to be safe is to check up on its recommendations.

For example, another thing I've been doing is working slowly through a text that's a bit beyond me (Una Historia de España by Arturo Pérez-Reverte) and it does a wonderful job of going line by line with its explanations—but I tend to follow up right after with a google search to see if there's divergence or if I'm on the wrong side of a hallucination.

[Cainntear]

Don't want to derail the thread What is the best advice you've been given? so there was something s_allard said that I thought I should bring into this thread in order to respond.

The tools and resources we have at our disposal now were unimaginable only 30 years ago. Add to this the AI revolution of which ChatGPT is just the beginning and we have a new dawn for language self-learning.

There was an exchange between me and sdw that kind of talks about the heart of the matter:

As far as I can see, ChatGPT is recommending something that's a common learner error.

I'm sure I've said something about "beber algo" and native speakers have recommended dropping the word "algo", because it's redundant.

"voy a beber" is effectively equivalent to "I'm going to go drinking" or "I'm going to go [out] for a drink" -- it's much more about the action and the social stuff than it is about the mechanical process of lifting a glass to your lips.

Ah that's great to know!

But it's a great point that GPT is a next token predictor so there are bound to be several of these kinds of mistakes. It does very well on some things but the only way to be safe is to check up on its recommendations.

As sdw said, ChatGPT is a token predictor -- it does a hugely complex statistical calculation to decide what the next word is, but crucially, it does this based on context: what is the most likely answer given to the question as set?

The problem is that in a lot of cases, questions set in language learner groups on social media will very often result in more incorrect answers than correct ones. As such, if the AI is asked a question in the style of a Facebook learner group question, there's a very high chance that it'll give a wrong answer.

I decided to dive into the "beber algo" thing a bit, and with dropping "algo", I couldn't replicate the incorrect behaviour I reported on sdw's screenshot -- i.e. it never once tried to "improve" voy a beber by suggesting voy a beber algo. The questions I'd used were "Is there anything wrong with this sentence?", "What is wrong with this sentence?" and "Spot the error in the following:" each followed by the same question: "Quieres ir a beber conmigo?".

The reason for this sequence of prompts (in different chat windows so one wouldn't affect the others!) is because I was trying to gradually increase the suggestion that this was wrong to see if ChatGPT would pick up on it and act as though it was wrong, but it kept insisting it was right, and didn't once suggest adding "algo" as a means of improving it.

My last attempt was even more extreme. I asked:

My teacher marked this as wrong. What would be the correct form?
Quieres ir a beber conmigo?

...and I got this reply:

It seems like your sentence is already in the correct form. "Quieres ir a beber conmigo?" is a grammatically correct way to ask "Do you want to go grab a drink with me?" in Spanish. If your teacher marked it wrong, you might want to check for any specific feedback they provided to understand the reason for the correction.

Even the act of saying the teacher had marked it wrong didn't lead to suggesting that I might add "algo".

I did the same experiment by adding in "algo", i.e.

My teacher marked this as wrong. What would be the correct form?
Quieres ir a beber algo conmigo?

And while the wording of the response was superficially different, the message was fundamentally the same; there was no attempt to change the sentence at all, although in this case removing "algo" would actually have improved it.

So the behaviour that sdw saw was very specific to the context provided, as it just doesn't seem to happen elsewhere.

Of course, I'm using free ChatGPT with GPT 3.5, and I can't rule out the idea that GPT4 might have "improved" in the sense that it now more accurately replicates the bad advice given by anglophones who incorrectly believe they're good at Spanish....