Using ChatGPT for Language Learning

This video shows examples of how well OpenAI’s ChatGPT might be used by people learning and teaching second languages. The video was recorded on December 7, 2022, a week after the release of ChatGPT. The speaker is Tom Gally (gally.net), a professor in the Centre for Global Communication Strategies at the University of Tokyo. Examples shown in the video are linked from the following page:
www.gally.net/temp/202212chat...
Introduction 0:00
Contents 0:58
Explaining meaning in context 3:00
Correcting and explaining grammar 13:31
Dictionary definitions and examples 16:37
Text annotation and translation 22:16
Vocabulary quizzes 26:26
Closing 28:48
Other videos in the same series:
“ChatGPT and the Future of Language Learning”
• ChatGPT and the Future...
“ChatGPT and Language Education”
• ChatGPT and Language E...

[@Tom_Gally_UTokyo]

A little over a month after I made this video, on January 16, 2023, Microsoft announced that application programming interfaces (APIs) for ChatGPT and other OpenAI programs would be accessible through their Azure cloud services. It will be interesting to see how developers use those APIs to make tools customized for language learning. Microsoft’s announcement is here: azure.microsoft.com/en-us/blog/general-availability-of-azure-openai-service-expands-access-to-large-advanced-ai-models-with-added-enterprise-benefits/ Then, on February 1, OpenAI announced that they would be rolling out a subscription-based version of ChatGPT with enhanced features while still keeping the free version available. That announcement is here: openai.com/blog/chatgpt-plus/

[@verbling.com.teachers.marie.]

I appreciate your content. I was able to take lesson content from an existing English Language student book and add specifics about the adult learner (job, business environment) to generate more realistic, engaging and relevant lessons for that individual learner. I disagree that ChatGPT will replace the critical thinking stimulation that language teachers introduce to language learning, the twists and turns in emotion during a conversation that occurs naturally without robotic prompts, turn taking, etc.

[@Tom_Gally_UTokyo]

@@verbling.com.teachers.marie. Many thanks for the reply and for the valuable comment. I hadn't thought of using ChatGPT to adapt existing ESL content for specific learners. That's a great idea. Thank you for sharing it. I agree with you that a human teacher is superior to ChatGPT for the critical-thinking and affective aspects of interactive learning-when such a teacher is available for one-on-one or small-group interaction. In many schools, though, because of budget limitations, languages have to be taught to large classes, and the teacher has little time to interact with each student individually. If the bots that are likely to be appearing soon have enough (simulated) personality to keep students engaged, then they are likely to be adopted for language practice even if they still seem a bit, well, robotic. As for whether these bots will eventually be able to simulate critical thinking and affective interaction well enough to replace humans even when a human teacher is available for one-to-one teaching, the jury is still out. As I mentioned in another video, I think humans will continue to play an irreplaceable role in motivation and as role models. As conversation partners and teachers, though, it might be difficult for us to compete with robots in terms of cost, 24-hour availability, and memory of the specific learner’s acquisition status.

[@mrstalder]

Thank you for all these examples. I have started using ChatGPT for my teaching and it's amazing how much time it will save me.

[@Tom_Gally_UTokyo]

Thank you! Please share your ideas and experience about teaching with ChatGPT as widely as possible. It’s still very new technology, and teachers don't yet know how to use it well. I hope that we will start hearing from independent language learners soon as well, as ChatGPT also seems to have a lot of potential for self-study.

[@nsreynolds8561]

I’m doing the same thing with my teaching. It is a game changer.

[@joshlee5294]

wonder how do u integrate it in ur teaching process?

[@nsreynolds8561]

Thank you for a fabulous video explaining some of the uses of ChatGPT in language learning. I teach English language learners (adults) in El Paso, Texas. I have been completely blown away with the capabilities of this newest technology. You have given me some wonderful new ideas on how to use ChatGPT language teaching and learning. I will be sharing your videos with my colleagues.

[@Tom_Gally_UTokyo]

Thank you for the comments! I feel as though I have just touched the surface. Please delve into the use of ChatGPT and similar software yourself and share what you learn as widely as possible.

[@nsreynolds8561]

@@Tom_Gally_UTokyo I’m doing just that. It is so much fun to see people’s reaction to this incredible technology.

[@Tom_Gally_UTokyo]

@@nsreynolds8561 Great!

[@joshlee5294]

how do u utilize this tool now, just out of curiosity

[@erosheve]

I personally find chatGPT very helpful at writing example topics for advanced language tests (B2 level)

[@Tom_Gally_UTokyo]

Yes. Its default writing mode seems to be just what test-makers like: logically organized and no language mistakes, though without much personality. It can write in many other ways, though, if you ask it to in the prompt.

[@YourFavoriteEnglishTeacher]

Thank you for these great suggestions for using chatgpt in our teaching! I really liked the sample sentences it generated, providing so much meaningful context, and I will encourage my university students to use this for vocabulary learning. To add to your list, Wwen asking for collocations it not only gave me various collocations for a word, but also provided sample sentences with them! It is great! 👩🏼‍🏫😁

[@Tom_Gally_UTokyo]

Thanks for the comments! There are probably many other ways to use ChatGPT in language learning, too. I'll be interested to see what you and others come up with.

[@telmo_a]

I naively assumed that, because ChatGPT is a "chatbot", it would be possible to practice some German simply by having an informal conversation with it. But it doesn't seem designed to keep meaningful, flowing conversations. For example, I told it I have a cat, and it just spewed out a bunch of information about having cats, instead of, say, asking me about my cat's name, or whether he is big or small. I asked it to ask me something about my cat, and it generated a list of questions. I asked it to ask me only one question, and it asked me how do I take care of my cat, then immediately flooded me with information about the importance of taking good care of cats. I asked it to ask me a question, then stop and wait for my answer, but it just repeated the same process: a question immediately followed by a load of info essentially answering its own question. It's like talking with someone who isn't truly interested in anything you have to say and just keeps showing off their own knowledge about everything. 🙂 I mean, it's impressive, just not designed to do what I hoped for.

[@Tom_Gally_UTokyo]

That's interesting. Were you giving it prompts in German or in English? I've been able to test it only in English and Japanese, and it responds more appropriately to prompts and requests in English than in Japanese. Though I have gotten unwanted Wikipedia-like explanations in English sometimes, too.

[@telmo_a]

My prompts were in German (though quite basic, A2/B1 German). However, in your latest video I noticed an example of conversation in English that looked a lot more like what I wanted. So I gave it another try, this time phrasing the German prompt more like in your example: "I have a cat. Ask me something about my cat." Surprisingly, not only the bot's answer was now a lot more concise, without any Wikipedia-like ramblings, the bot also understood that I wanted it to keep asking me questions on the same subject! So, "conversational" performance appears to depend a lot on the way prompts are given. Many thanks for your answer and videos. I'll certainly keep exploring the use of this bot as a tool for practising German.

[@Tom_Gally_UTokyo]

@@telmo_a Thanks for the reply! So it does seem to work okay in German when prompted properly. That's good to know. I suspect that, once OpenAI makes APIs available to developers of language-learning software, it will become possible for learners to choose from a menu of preset interactions that are known to work rather than the learners having to come up with their own prompts.

[@SuperZardo]

You need to tune your task, the chatbot needs carefull tuning if you want him to do a complex task. Describe the steps, if necessary. If it does something you do not want him to do, tell him not to do it. By the way, do not expect that it will take any particular interest in your cat or yourself, the conversation will therefore never be personal.

[@ITNoetic]

You can do back and forth role play conversations, but you have to say that that's what you want. You can't just say "I have a cat" and expect it to know that you want a conversation and not cat facts

[@academy-t1]

to me, if teachers and tutors of English can use this software effectively, it is to contribute to their businesses. I think it is not something like an alternative for teachers, but naturally teachers and tutors have to improve themselves, and I think this is a perfect tool for English teachers. :) thanks for the video.

[@Tom_Gally_UTokyo]

Thanks for the comment! I also hope that teachers and tutors will be able to incorporate this kind of software effectively in their teaching.

[@annann4863]

Totally agree with this.

[@joeandcoffee]

I am a language tutor, and I have adopted this technology into my classroom

[@joshlee5294]

@@joeandcoffeemay i know in what scenario would u use it

[@erosheve]

Thank you so much for a great test and fresh ideas of how one can use this truly remarkable piece of software! 🙏

[@Tom_Gally_UTokyo]

Thank you for your comments. I feel like we have only just begun to figure out how ChatGPT can be used by language learners. I hope that you and others experiment and explore and share what you find widely.

[@dogchaser520]

For teachers: this could be used to generate nearly infinite worksheets. For example, tell it to generate sentences using these vocabulary words and/or this grammar point, and tell it where you'd like to include blanks (for the subject, for the verb, etc.), and it can do it. I've also been able to coax it into breaking down stories into very simple structures, such as SVC. This could help students understand the meaning of some more complex structures by inference. It's a shame it "hallucinates" some grammar explanations, though. I think if you keep things very simple, it will usually come out correct ... but it's difficult as a language learner not being able to trust it! I still plan to use it in my own language learning.

[@Tom_Gally_UTokyo]

Thanks for the comments. Those are good ideas. I agree that the grammar “hallucinations” are a definite drawback for application to language learning. They seem related to ChatGPT’s mediocre mathematical ability, which might be due to its reliance on analogy rather than analysis to solve abstract problems. I hope that the problem will be solved eventually.

[@dogchaser520]

@@Tom_Gally_UTokyo I'll be willing to bet that in some time, a person will try fine-tuning it to language learning. Providing a fairly comprehensive grammar textbook might be enough to overcome most of it - and OpenAI does allow a person to train models on whatever you'd like. Creating good training data can be a challenge, though. The overall problem is that it's a predictive text model that has no actual understanding of anything that it generates. It just uses a very large collection of text and extrapolates what is the most likely word to come next based on the prior context, with some random variance based on settings. That's an oversimplification, but for the most part it's true. The ChatGPT model doesn't allow you to modify settings common in most text-generation AI, such as Top P, temperature, etc. I'm willing to bet that lowering Top P (so that it selects only the most likely of words to come based on past context ... a value of 0.4 meaning it only chooses from the words which have a combined probability of 40% of being next) would eliminate some of the issues, but it would need testing. It also would sound much more robotic. But for a grammar explanation, that might be what's needed. I'm betting ChatGPT uses a middle range value for temperature, Top P or both, so as to be able to stay dynamic and respond to a variety of questions. Adjusting values on the fly based on what a person needs for a given question is probably in the works for future AI interfaces!

[@Tom_Gally_UTokyo]

@@dogchaser520 Interesting comments. Thank you. I’ve seen other people also describe ChatGPT as a “predictive text model” that “extrapolates what is the most likely word to come next based on the prior context,” but I’m not sure how adequate that explanation is for transformer-based software like ChatGPT [1]. The “self-attention” mechanism seems to make it fundamentally different from a word-by-word, autocomplete-like device. To give an example: Last month, I posted a comment to Hacker News in which, in passing, I showed how ChatGPT can explain the meaning of a nonce word that had been used only once in that meaning in human history (probably) [2]. The ability to explain the meaning of a previously unattested word based its context cannot, I think, be adequately explained as a predictive text model. At least, that explanation is not very convincing to me. [1] en.wikipedia.org/wiki/Transformer_(machine_learning_model) [2] news.ycombinator.com/item?id=34556855

[@dogchaser520]

​@@Tom_Gally_UTokyo That's a wonderful example and edge case. You're right that GPT is not as simple as what we've seen with, say, autocomplete. I had to read up about transformer models and attention. I'm not an AI researcher but did study computer science (very long ago). From what I understand, the transformer model does still rely on probability of the next token to be generated, but uses attention as a way to further refine those probabilities by focusing on what's most important in the text. For the specific example of "flammilingus," I think there are some hints for it to draw on. For one, GPT's internalized what are probably thousands of portmanteaus, and its vast dataset might allow it to respond about them in a way that's meaningful. (A good test of this might be to invent a few and ask it what it thinks they mean.) For this specific word, though, there are mentions in a blog (The Thinking Housewife) and a book ("B & Me: A True Story of Literary Arousal") available online, and those contexts alone might provide enough of a "framework" (in a loose sense of the word, not implying any understanding) for GPT to give a reasonable answer. The last bit that really convinces me it's still generating bit by bit after all its computation is when using the Playground in OpenAI and turning on the "Show probabilities" option. This shows the likelihood of the next word to be generated. Though that's simplifying it for us, as it operates on character chunks/tokens rather than full words. Here's an example of testing out an imaginary portmanteau. Not the absolute best (after thinking I'd made it up, I found there are examples on the web, but none seem to be similar to what it came up with). I still think it shows it's able to puzzle it out without an existing example: Me: This word doesn't exist, but if it did, what do you think "pornographite" would mean? GPT: Pornographite would likely be a type of mineral or rock, similar to graphite, which has sexual undertones or themes in its coloration and structure. (I had to tell it the word didn't exist first in order to get it to speculate. Otherwise you have to nudge it a bit after.)

[@Tom_Gally_UTokyo]

@@dogchaser520 Thanks for those comments, too. “Pornographite” is a nice one. Now that OpenAI has released APIs, I hope to find some time to test it more systematically. But I did just try my own portmanteau prompt, shown below, with the regular ChatGPT. The first three sentences in the target paragraph come from an essay by John Updike; the last sentence with the portmaneau is my own creation. (John Updike would never have been so gauche.) In my test, it took some effort to get ChatGPT to recognize that “-dom-” was supposed to represent “dome,” the otherwise it was able to identify the meaning of the unknown word well. You might want to try this prompt and see what you get. Prompt: The following is the first paragraph of an essay by an American writer. The word “equaldomity” seems to have been invented by the writer, as I cannot find it in any dictionaries or on the Internet. Can you speculate about its intended meaning and etymology? On this Sunday morning, all over Chicago, churchgoers are settling to hear a sermon and to sing the praises of the Lord; let us, then, in synchrony sing the praises of freedom and equality, those two bluebirds of hope and aspiration swooping in our American skies. The two concepts nest at the heart of what we like to think makes the United States an exemplary and revolutionary place; yet neither is unambiguous or without its problematical aspects. Equality, I would say, is a practical concept, and freedom a subjective, rather metaphysical one. The two intersect in something one might call “equaldomity,” an abstract edifice like the physical building where Congress meets in Washington, with a magnificent central peak winged by the two contrasting concepts.

[@SeanTechStories]

Love this deep dive video!

[@Natalia_English_Teacher]

Also, a great idea that works well for me is creating dialogues on a certain topic and then making a quiz out of them. "Write a dialogue for a young learner of English between Mary and George at A1 level of English about chores and shopping using the constructions: as...as, not as...as, enough and not enough" Now, make a quiz out of this dialogue. Works great.

[@Tom_Gally_UTokyo]

Interesting! I encourage you and other language teachers to share your ideas as widely as possible. It’s a brand new tool, and we’re just beginning to learn how to use it.

[@MusicReaction]

GIGO can be applied here to the tasks you assigned to GPT, if you redefine the parameters to your request it will give a better response, for example 'oversight' would've been explained better if you'd have defined the level of language the learner is at, ie B1 or pre-intermediate... Also articles are an extremely confusing item of grammar, I have trouble explaining it, and many attempted explanations of it on the internet are erroneous, so the samples GPT learned from would've been confusing, hence the mistaken response. Overall I really appreciate your lecture here, it will help me to get a grip on AI quicker and teach IELTS better. All users just have to be careful how they word requests.

[@Tom_Gally_UTokyo]

Good points. Thanks.

[@Natalia_English_Teacher]

Thanks for the video, it was useful. My son showed me ChatGPT and inspired me to use it in my work as an ESL teacher. Now I'm thinking of how to make profit out of it)))

[@Tom_Gally_UTokyo]

Good luck!

[@the-language-learner]

Hello Tom, I wish I had watched your video before uploading my own video about ChatGPT. You brought up requests I hadn't thought about, such as explaining meaning in context. The possibilities with this new tool are endless. I also demonstrated in my video that it can't answer everything. For example, when I asked how many hours of study it would take for a 70-year-old Japanese person who can't read or write to reach B1 in French, it couldn't answer the question. Let's see if ChatGPT can handle this question and also make quizzes as you experimented with in a couple of months. I would say it has the potential to." English is not my first language and the previous sentences have been corrected using chat GPT. ;)

[@Tom_Gally_UTokyo]

Thanks for the comment. Yes, ChatGPT is likely to get better over time. Also, it seems that one can get better results sometimes by adjusting the prompts. For example, it often cannot solve simple mathematical or logical word problems correctly when asked only for the solutions; if it is asked to explain each step in its reasoning, though, it solves the puzzle correctly more often (though still not always). I was able to get better vocabulary quizzes than in the video by using this prompt: “Create a multiple-choice vocabulary quiz with ten questions for an intermediate learner of English. Make sure that only one of the choices has the same or similar meaning to the word being asked about.” However, I haven't yet figured out how to adjust the prompts to get it to give correct grammatical explanations.

[@jackjhmc820]

Anyone tried writing japanese and see if chatgpt could correct the japanese grammar? There s no grammarly in japanese.

[@Tom_Gally_UTokyo]

In the tests I've done, it can correct learner’s mistakes in Japanese and make the text more natural, but its explanations of its corrections are often unhelpful or incorrect. (Its explanations of English grammar are often wrong, too.) Below is a test I did just now. The text came from an online corpus of texts written by learners of Japanese (corpus.icjs.jp/corpus_ja/getCorpus.php?corpusId=564&errorId=2). To my eye, ChatGPT's rewritten version does look much better. But some of its explanations that follow, such as “incorrect verb conjugation” and “where the speaker is expressing a general trend or observation, using the more formal 「になっています」 is more appropriate” are, I think, wrong. (Its comment about “appropriate and accurate vocabulary and grammar” is just the sort of unhelpful help provided by native-speaking teachers who are unable to explain their linguistic intuitions. I’ve done that many times myself😆) In any case, I suggest you try asking ChatGPT to help you with your Japanese and see if its responses seem helpful and correct. The quality of the results may vary a lot depending on the type of texts and the prompts that you use. Here's my test: Me: The following paragraph was written by a nonnative speaker of Japanese. Produce a new version of this in Japanese with the grammar and vocabulary errors corrected. 最近、的確なリサイクルをするのは大切になっている。リサイクルとはゴミを別れる方法のことだ。特に、紙、プラスチック、缶などを別れる方法のことだ。この方法が考えられたのは地球を守るからである。現在、様々なことが行われている。よく知れらた例は地球温暖化のことだ。とはいえ、同様に重要なことは、ゴミが不適切に捨てたら、動物が痛める。多くの写真のはゴミに引っ掛かる鳥や亀のである。そのため、悲しいより地球の生態系が壊れるようになりかねない。それをしたりすると、地球で活きにくくなる。これから私たちがしなければならないことは適切にゴミを捨て、別れることである。確かに、いつもゴミをゴミ箱にすてなければならないが、加えて缶やペットボトルなどを適当は箱にすてることになっている。その後、動物が見つけられない場所に置くと思う。そのように動物が傷つけないのである。これは地球温暖化のとまり方か分からないが、いい始めだと思う。 ChatGPT: 最近、リサイクルを適切に行うことが重要になっています。リサイクルとは、ゴミを分別する方法のことです。特に、紙、プラスチック、缶などを分別することを指します。この方法が考えられた理由は、地球を守るためです。現在、多くのことが行われていますが、代表的な例は、地球温暖化の問題です。一方で、不適切なゴミの処理によって動物が被害を受けることも重要な問題です。多くの写真では、ゴミにかかった鳥や亀が写っています。そのため、不適切なゴミ処理によって、生態系が壊れてしまう恐れがあります。このような状況が続けば、地球上で生活するのが困難になります。これから、適切にゴミを分別し、処理することが私たちの責務です。例えば、通常はゴミ箱に入れますが、缶やペットボトルなどは適切な箱に入れることが求められます。また、動物が触れない場所に置くよう心がけましょう。このようにすることで、動物に被害を与えないことができます。地球温暖化の問題を完全に解決することは分かりませんが、このような取り組みは良いスタートだと考えます。 Me: Why did you change the sentence 「最近、的確なリサイクルをするのは大切になっている。」 to 「最近、リサイクルを適切に行うことが重要になっています。」? ChatGPT: I changed the sentence because it contained grammar errors, such as incorrect verb conjugation, as well as vocabulary errors. The corrected sentence uses more appropriate and accurate vocabulary and grammar, making the meaning clearer and easier to understand for a native Japanese speaker. Me: Specifically, why did you change 「大切になっている」 to 「重要になっています」? ChatGPT:「大切になっている」 translates to "has become important" which is grammatically correct, but "重要になっています" is a more natural and commonly used expression to express the same meaning, making it a clearer and more suitable choice in this context. Me: In that sentence, why did you change になっている to になっています? ChatGPT: 「になっている」 is the informal present tense of "to become". However, in this context, where the speaker is expressing a general trend or observation, using the more formal 「になっています」 is more appropriate. Additionally, the use of ます in the sentence makes it sound more polite and less casual.

[@dariuszjaneczko1927]

This is only a sample. It can create texts, tests and lists on various topics, amounts of words, grammar, and various levels from a1 c2 for sure

[@francisjacquart9618]

THANKS A LOT DEAR SIR FOR THIS VERY INTERESTING VIDEO AND ALL THE EXPERIMENTAL TRIALS. IT IS EXTREMELY INSPIRING AND WHEN THIS NEW TOOL WILL BE HONED TO PERFECTION, THE SKY WILL REALLY BE THE LIMIT AND IT ALREADY IS! LOVE AND TAKE CARE!

[@Tom_Gally_UTokyo]

Many thanks for the kind comments!

[@danielbailey1855]

Thank you for making this video.

[@gabrielbour1722]

der Tom I tried this prompt for quiz and had a better result : create a vocabulary quiz for an A1 level student What is the word for a building where people live? House, car , street, tree. What is the word for a piece of paper with writing or printing on it? Book, pen newspaper chair.

[@Tom_Gally_UTokyo]

That's interesting! Thanks for letting me know. I tried your prompt just now, and it worked for me, too.

[@Nighteye88]

I noticed that it with Japanese constantly does the "he she he she" examples too. Overusing it a bit I think. I've only been using it a few days for Japanese but it's been awesome for the most part even with the mistakes it's so nice to get examples sentences in seconds instead of searching the internet for each one.

[@Tom_Gally_UTokyo]

Yes. It should be even easier to use after it gets integrated with other software. Imagine being able to click on a word in a text you're reading and immediately get not only a definition but also five example sentences. And, unlike current dictionary look-up software, it will be able to identify the sense of the word in that context and explain the meanings of words and phrases that don't appear in human-created dictionaries. I just tried asking ChatGPT to explain a phrase in some informal unedited text, and it did perfectly: Me: The following comment and reply appeared on the site Hacker News. What does the sentence “I think this is one of the better, well my dad can beat up your dad type of stories” mean? Comment: My father Bryce Bayer studied this question fifty years ago at Eastman Kodak; his approach is known as ordered dithering. One is effectively posterizing a grayscale image, and his primary goal was to reduce artifacts drawing unwanted attention to the borders between poster levels. With improvements in hardware other approaches to dithering took over. The last time I saw my father's pattern was on a DEC box logo. He moved to the color group at Kodak, and designed the "Bayer filter" used in digital cameras. Reply: I think this is one of the better, well my dad can beat up your dad type of stories. I know you absolutely didn't mean it that way, but there's a part of me that read it that on the second reading of it. Oh yeah, well my dad...! kind of thing =) ChatGPT: The sentence "I think this is one of the better, well my dad can beat up your dad type of stories" means that the reply is saying that the comment is a good story and is making a comparison to a type of boastful statement where someone boasts about their father being better or stronger than someone else's father. The speaker is making a joke about how the comment reminded them of this type of boastful statement and is not taking the comment seriously. Me: Can you give some other examples of “my dad can beat up your dad” being used in natural contexts? ChatGPT: Some examples of "my dad can beat up your dad" being used in natural contexts are: "Don't mess with me, my dad can beat up your dad." "My dad is a professional boxer, so don't even think about starting a fight with me, my dad can beat up your dad." "I may be small, but don't underestimate me, my dad can beat up your dad." "I'm not afraid of you, my dad can beat up your dad." "Don't brag to me about your father's strength, my dad can beat up your dad." It's worth noting that this type of statement is often used in a playful or joking manner, and not meant to be taken literally.

[@dogchaser520]

You can specifically tell it to use a variety of subjects, or more complex sentences. Or try using the Playground in OpenAI for a broader experience.

[@AspireNation]

thank you for this video, it's really helpful.

[@robertolanzone]

I've been trying to use it to learn Japanese but time and time again it spouts some mistakes and I was lucky enough to notice them (but who knows if there were more I was unaware of?) Even worse, after letting it know there is a mistake, it apologized and gave a whole different still incorrect answer. It didn't realize what the actual answer was until I spelt it out for it, and this happened more than once. I was really hyped but this completely turned off my interest: I can't rely on a teacher that teaches me wrong pronunciations and/or translations every now and then

[@Tom_Gally_UTokyo]

Thanks for the comment. Have you been using ChatGPT as a conversation partner, or have you been asking it to teach you aspects of Japanese grammar explicitly? In my tests, though it seems somewhat less fluent and natural in Japanese than in English, I have yet to see any grammatical or vocabulary mistakes in the Japanese it produces, and I think it should be possible to practice conversations with it. However, its explicit explanations of Japanese grammar have often been wrong, as have its explanations of English grammar. Just as it doesn't seem to be able to do quantitative reasoning very well, it also doesn't (yet) seem able to reason explicity about language very well. But its production of language is amazingly good. Also, as I show in the video, it is quite good at being able to explain what a word means in a particular context. In some later tests, I have asked it to guess what the meaning of an obscure word might be based on its context, and it is quite good at that, too-better than the average native speaker, probably. Below is a dialogue I had with it yesterday. I hadn't known the slang use of “G” to mean "gokiburi," and ChatGPT didn't seem to know it, either. With some hints from me, though, it was able to guess it. Me: 次のパラグラフは日本人作家のエッセーからの抜粋です。夏の「G」の意味がわかりますか。 僕が大学に行ったのは27歳の時だ。やりたいことがたくさんあって高校を出てから約10年間、自由奔放に生きた。海外に行くため長期で働けないから30種くらいアルバイトを転々とした。風呂なし共同トイレのぼろアパートで極貧生活を味わったこともある。安いパンで飢えを凌ぎ、金欠で銭湯に行けず、ガスがきていなかったため冷水で髪を洗い体を拭いた。夏は「G」が大量発生し、冬は凍えかけた。与えられた環境の外部は予想外の出来事に満ちていた。 ChatGPT: 「G」は虫のことを指している可能性が高いです。夏には虫が多く発生し、作者はそれによって苦しんだと書いています。 Me: 日本語で「G」で始まる虫のようですね。なんでしょう。 ChatGPT: 「G」で始まる虫については、詳しい情報がありませんが、「G」で始まる虫名として、「ゴキブリ」があります。これは、人間の生活に深く関わっている虫で、夏に特に多く発生します。それが、作者の記述に基づいている可能性があります。

[@mc101]

I just used this prompt... and it worked GREAT: Me: You are to play the role of my good natured German instructor. You are to follow two rules: Rule 1 - You will always correct my German if I make an error and explain what I said wrong. Rule 2 - You will always reply first in German then second in English. If you understand, then please repeat these two rules while applying both rules. ChatGPT: Ja, ich verstehe. Regel eins besagt, dass ich immer Ihren deutschen Fehler korrigiere und erkläre, was Sie falsch gesagt haben. Regel zwei besagt, dass ich immer zuerst auf Deutsch antworte und dann auf Englisch. Wenn Sie möchten, können Sie beginnen. Yes, I understand. Rule one states that I will always correct your German errors and explain what you said wrong. Rule two states that I will always reply first in German and then in English. If you'd like, you can start.

[@Tom_Gally_UTokyo]

Thanks for that idea! When I made this video back in December, it had been only a week since ChatGPT was released and it didn’t yet occur to me to use long prompts like that. I just tried your prompt myself with GPT-4, changing German to Japanese and making the second rule “You will always reply first in Japanese standard script, second the same Japanese in romaji, and then third in English.” That worked okay and it was able to correct mistakes. Its feedback on the mistakes was not always ideal, though. For example, when I omitted from the past tense of a verb the character that indicates a long consonant-a common mistake for learners of Japanese-rather than inserting that missing character it gave me a different, more polite form of the verb. If I were an actual learner, I probably wouldn’t have understood what my mistake had been. After I added “When I make a mistake with long or short vowels or consonants, please point out those mistakes rather than giving a different form of the word.” to the prompt, it was able to do better. Correcting a learner’s mistakes can be challenging even for experienced human teachers, as it can be difficult to know what caused the learner to make the mistake, whether the learner is ready to learn from a correction of the mistake, and, sometimes, what the learner was intending to say in the first place. It will be interesting to see whether actual learners who have interacted extensively with both human teachers (experienced or inexperienced) and GPTs will prefer the humans or the AI for grammatical instruction and correction.

[@mc101]

@@Tom_Gally_UTokyo I'm glad my idea has some benefit. Thank you for your feedback. Keep up the great work.

[@nasirabbasi6060]

Thanks for sharing.

[@Tom_Gally_UTokyo]

Thanks for watching!

[@nasirabbasi6060]

I'm an EFL teacher and I'm also interested to learn about it. Regards

[@SpotlightGraphics]

Imagine combing ChatGPT with Obsidian md. I feel like you can learn a lot very fast.

[@Tom_Gally_UTokyo]

Good point. It will be interesting to see how ChatGPT and similar software are integrated with language-learning software. I expect that we will start seeing examples within months. Imagine if you had a language teacher who remembers all of the vocabulary and grammar that you have been exposed to and where your strengths and weakness are, and can then devise, on the spur of the moment, interactive practice materials tailored to your needs and interests. With the suitable interface, ChatGPT seems capable of doing that.

[@wjdouma734]

このAIすごいね、もちろん微調整する必要があるけど、学習者に役に立つかもしれないね。 特に新しい語彙を暗記する方法のアドバイス好きでした。 将来的にどうなるか気になります。

[@Tom_Gally_UTokyo]

私も気になって仕方がありません。 コメント、どうもありがとうございました。

[@bokshil]

Also Arabic Very week level as a native speaker it was below average

[@mrLeonidaszazelis]

I must point out that you forgot to say that we have to train ChatGPT to do things, not only expect it to provide us all we need.

[@Tom_Gally_UTokyo]

Thanks for the comment. Yes, the wording of the prompts is important. At the time I made the video, the only way of interfacing with ChatGPT was through the browser interface, and it did not yet remember discussions between sessions, so that prompting could be awkward, repetitive, and time-consuming. Now that it can be interfaced through APIs and longer prompts can be fed to it automatically, it should be possible to create user interfaces that are better tailored to the needs of language learners. I won’t be developing such interfaces myself, so I look forward to seeing what other people develop.

[@jk35260]

Teachers offer a variety of roles in education that ChatGPT is not able to perform, such as providing discipline, implementing a curriculum, facilitating interaction among students, and so on. So teachers are not going to lose their job.

[@Tom_Gally_UTokyo]

Thanks for the comment. I agree about those other roles. I mention some of them at the end of my “ChatGPT and the Future of Language Learning” video. I don't know, though, whether all teachers, or all educational systems, will be able to shift their focus to the roles best performed by humans. Educational administrators in particular may find the lower cost of AI sufficient reason to reduce the number of human teachers.

[@jk35260]

@@Tom_Gally_UTokyo ChatGPT will revolutionize the manner in which individuals acquire knowledge across a wide range of subjects. I think students will hold a high regard for educators who possess the skills to seamlessly integrate ChatGPT in their language learning process, thereby augmenting the pace and quality of their learning.

[@gwendyr8140]

Naaah, i'm not the first, who want to use it like this?

[@henrytsao1]

Thanks to GPT, no more grammarly and its constant yellow line of telling you how shit your writing is.

[@suusususususussuussusu3392]

Hi, professor Gally. Whenever ChatGPT stops, you just have to write something like ''go on''' and it will continue with the response. Also, I made the very same question about the ''a house'' change and it gave me the correct answer instead (maybe an update). I'm learning english so please tell me if I misspelled something. THANKS!

[@Tom_Gally_UTokyo]

Thank you for your reply! Yes, after I made that video I learned that you can get ChatGPT to continue responses by asking it to "go on." "Continue" also works. Your comment looks great. I hope you continue to enjoy learning English and that tools like ChatGPT will help!