Using ChatGPT as a Language Tutor

[Le Baron]

It seems to me it would just be easier to just go on a forum and ask a Spanish person. And if the objection was that this might also end up supplying the questioner with a subjective answer, at least on a forum there would be varied input from varied experience allowing it to show a rounded view (e.g. 80-90% concurrence upon a certain answer).

Perhaps someone should do an experiment using Chat GPT for 2-3 months in certain ways and testing it with a pre- and post Dialang test. No other resources, just GPT. To see if this noticeably improves: grammar understanding, writing, reading, speech (formal and informal), listening etc.
The questions I'd hope to answer would be: did GPT do something other resources can't? Was it quicker/more efficient (or the opposite)? Was the input effort less, the same or more?

All that would pique my curiosity more than supposition about what Chat GPT can 'supposedly' do. I think it's fun interacting with the Chatbot when I've done it, but if I knew it was not really adding much (vs the appearance of it) I'd never do it again. If the opposite I'd do it more.

[golyplot]

The interactive chatGPT only processes about one or two pages worth of material at a time. You can do a whole book but it will take some time, as you will be copying and pasting a lot. If you need to automate larger tasks you should write a script to use the OpenAI API (which I haven't used myself so I can't go into more details).

The other problem I ran into when I tried to do this in the past is that the output would quickly cause the prompt to fall out of the context window and instead of warning you, it would just seamlessly continue by hallucinating a continuation of the story instead of translating.

[MorkTheFiddle]

Using ChatGPT to generate questions about a passage.
emk, Bombobuffoon, and I have been volleying back and forth about using AI and language learning.
To recap what I have been saying, I am playing with the strategy of using Free Recall and then a follow-up test for consolidation.
Preparing the test yourself seems to sabotage the strategy, since you would have to read or watch or read and watch before the task.
But questions prepared by another person or by an AI free you from contaminating your knowledge.

As a test case, I used ChatGPT 3.5 and (later) the free version of Claude, to ask five questions about the Ancient Greek speech Philippic 3 by the Athenian orator Demosthenes. The speech was new enough to me. I had glanced at it several years ago, but not enough to begin to understand it. And The though the passage is in ancient Greek, the questions are in English and I believe they can be studied for their general appropriateness.

ChapGPT suggested a passage of 500 words in length to feed it. So using the text in Perseus, I carved out the first 500 words or so of the speech, fed the speech to ChatGPT, and awaited its questions.

Here are ChatGPT’s questions, and then five questions from Claude.AI, which I asked afterwards.
1. What is the main concern Demosthenes expresses regarding the Athenian assembly?
2. According to Demosthenes, why does Philip of Macedon not only wrong the Athenians but also others?
3. How does Demosthenes characterize those who prioritize flattery over speaking the truth?
4. What does Demosthenes argue about the concept of freedom of speech in Athens?
5. What does Demosthenes suggest is the consequence of the Athenians' current attitude towards Philip of Macedon?
After reading the passage, I could then give an adequate answer to only the first question.
Here are 5 questions from Claude.ai free version:
1. What is Demosthenes' main accusation against the Athenian politicians at the beginning of the passage?
2. What does Demosthenes criticize regarding the behavior of certain orators in the assemblies of the people?
3. What paradoxical contradiction does Demosthenes point out between the freedom of speech in the assemblies and the lack of freedom in giving advice?
4. What does Demosthenes say has happened to the Athenians because of this situation in the assemblies and the councils?
5. Despite the serious problems, what hope does Demosthenes offer at the end of the passage?
After reading, only question 5 got an adequate response from me.
Conclusion: The questions and my lack of answers do show I did not really understand the passage.

Anyone interested can find the passage and an associated English translation here:
Text from Perseus, Demosthenes Philippic 3, https://tinyurl.com/ysbx59bu
or, long version: https://www.perseus.tufts.edu/hopper/te ... ection%3D1

[emk]

Here are a number of things I'm doing with GPT-3.5-Turbo and GPT-4. I'm mostly using it at the API level, not at the chat level.

• Translation. GPT-3.5-Turbo is surprisingly good at translating short passages (a few sentences or a paragraph) from French or Spanish to English. I make heavy use of it for subtitle and song translation, and it produces perfectly adequate results. For particularly demanding texts, I've also tried GPT-4, which competes well with DeepL and Google Translate. However, GPT-4 is a complete and unmitigated disaster at translating Middle Egyptian.
• Explanation of words in context. This is surprisingly good, though not perfectly accurate. I take sentence cards, mark intresting expressions with "[[ ]]", and ask GPT-3.5 to break down the phrase and explain it.
• Transliteration. This is actually done with OpenAI's other model, Whisper-1. It produces suprisingly good transcripts for TV shows with clear enuniciation, and acceptable ones for songs with relatively clear vocals. More difficult TV shows and movies produce mixed resutls.

Importantly, none of these tasks require 100% accuracy. I can deal with the occasional bad subtitle, or translation mistake, or incorrect explanation.

These models can also be fantastic for generating sentences and stories of various sorts. But I really wouldn't rely on them to correct user errors. If you do insist on trying error correction, definitely pay for GPT 4. It will at least sometimes tell you you're making a mistake, though not reliably so. GPT-3.5 is basically unable to recognize mistakes, either yours or its own, even if you explain those mistakes carefully.

[Carl]

If you do insist on trying error correction, definitely pay for GPT 4.

Or use GPT 4 or GPT 4-Turbo for free!

The costs of ChatGPT have come down rapidly. I paid US$20 for a month of ChatGPT Plus last year, primarily so I could try out the speech mode in the iPhone app. Now the speech mode is included for free in the app, albeit still with ChatGPT 3.5.

However, Microsoft now offers both GPT 4 and GPT-4 Turbo at no cost, in Copilot:
https://www.cnet.com/tech/computing/microsoft-copilot-now-has-gpt-4-turbo-for-free-what-to-know/

I use Copilot through Skype, which is the app I use for most of my language exchanges. There, my top contacts include my human language exchange partners plus Copilot.