Using ChatGPT to write Mini-Stories

Open Forum
1

I have been playing with using ChatGPT to write mini-stories, or mini-story like content, in some new languages I have been studying, and after Steve’s short on YouTube yesterday I thought I’d share the prompt I have been using to get pretty good results.

Prompt
Hi Chatty, can you write a simple story in Low Saxon that is 10 sentences long. This story should be told twice. Once from the perspective of the main character, and once from a third party perspective.

Can this be followed by a question and answer section of 8 questions each enumerated. The each enumerated question should be formulated like this:
#. Sentence from or about the story.
A question that reformulates the sentence from or about the story.
The answer to the question.

Chatty’s Response
Joa, ik kann di dat gern maaken. Hier is de Geschicht:

First-person perspective:
Ik sitt in mien Stuv un kiek na buten. De Sunne schient un de Vagels singt. Ik heff mi völlig entspannt. Ik denk, ik mook mi heüt mal een schöön Dag. Also heff ik beslaten, dat ik mi en lecker Stück Koken maaken will. Ik gaoh na de Küken un kieke, wat ik noch allens heff. Dann finn ik en Rezept för en bannig gode Kuchen. Ik goah an’t Wark un na en poor Stünn is mien Koken klaar. Nu sitt ik hier mit mien Kaffee un mien lecker Koken. Dat is doch mal wat.

Third-person perspective:
An een schöön Dag sitt een Person in ehr Stuv un kiekt na buten. De Sunne schient un de Vagels singt. Se is völlig entspannt. De Person denkt, se will sik heüt mal een schöön Dag maken. Also hett se beslaten, dat se sik en lecker Stück Koken maken will. Se geiht na de Küken un kiekt, wat se noch allens hett. Dann finnt se en Rezept för en bannig gode Kuchen. Se gaht an’t Wark un na en poor Stünn is ehr Koken klaar. Nu sitt se hier mit ehr Kaffee un ehr lecker Koken. Dat is doch mal wat.
Questions and answers:

  1. Ik gaoh na de Küken un kieke, wat ik noch allens heff. Wo geiht de Person hen, um to kieken, wat se noch hett? (Where does the person go to check what they have?)
    De Person geiht na de Küken. (The person goes to the kitchen.)

  2. Dann finn ik en Rezept för en bannig gode Kuchen. Wat hett de Person funnen? (What did the person find?)
    De Person hett en Rezept för en bannig gode Kuchen funnen. (The person found a recipe for a very good cake.)

  3. Se gaht an’t Wark un na en poor Stünn is ehr Koken klaar. Wo lange daurt dat, bit dat Koken klaar is? (How long does it take until the cooking is done?)
    Na en poor Stünn is ehr Koken klaar. (It takes a few hours until the cooking is done.)

  4. Nu sitt ik hier mit mien Kaffee un mien lecker Koken. Wat hett de Person, as se klaar is mit Koken? (What does the person have when they are done cooking?)
    De Person hett mien Kaffee un mien lecker Koken. (The person has their coffee and their delicious cake.)

  5. Se is völlig entspannt. Is de Person völlig entspannt? (Is the person completely relaxed?)
    Ja, se is völlig entspannt. (Yes, the person is completely relaxed.)

  6. De Sunne schient un de Vagels singt. Wat is dat Wedder? (What is the weather?)
    De Sunne schient un de Vagels singt. (The sun is shining and the

When Chatty’s stories get too long it cuts off in the middle of the Q&A like it did here :), but otherwise I have a hard time not being a little impressed with the results.

12 Likes
2

Ik kan ok plat verstoon. Mine Groseldern het dat freuer sproken. Dat de robot ok spreken künn, mokt mie spraklos. Kenn je dat verstoon. Nun in hochdeutsch: Liebe Grüsse aus Deutschland/Niedersachsen (Lower Saxony)

2 Likes
3

Ik kann Platt verstoon, man op Platt schrieven oder snacken is mi swaar. Ik heff en Paar Böker leest un BremenEins hett de Narichten op Platt. Deit mi leevt in Amerika, un ik kenn man een Persoon, de Platt snacken kann.

Liebe Grüsse aus Minnesota :).

3 Likes
4

Not mini-story related per se, but I see a use case for GPT in language preservation. I have used Chat-GPT to make similar content in:

  • Low Saxon/German
  • West Frisian
  • Northern Sami
  • Faroese
    It refuses in Greenlandic :).

In each case, my knowledge of the language is limited. I may have high comprehension, but don’t speak them. That said, the results look really promising, especially if this is the beginning.

2 Likes
5

Really cool! Since I cannot judge the quality at the linguistic level, let me just rant a little and express my continued perplexity about these language models [1].
I’m assuming that there is extremely little content on the Internet in Low Saxon and further, that whatever exists is only partially represented in the training data. So, the question poses itself, how is the model able to convincingly produce content in this language?
The answer must be related to the statistical nature of the model, my best guess is that it transfers knowledge from other, closely related languages. Or that it has abstracted certain fundamental concepts of Germanic languages, like grammar or etymology, and is able to ‘build’ a language even with very limited training data.

I’m not an expert but I’ve seem similar things with Whisper, the speech-to-text model, it uses the transformer architecture as well and is made by the same company, it’s obviously infinitely more primitive. Still, it has this uncanny ability to translate into languages other than English (which it is not trained to do) so I wouldn’t be surprised if similar things are going on here, just at a much higher level.

And another thing that is puzzling, does it internally translate? The prompt is in English after all. I remember that Openai states, the results may improve in translation tasks if the prompt is given in the target language [2].

Anyway, fascinating stuff, wish I had an actual understanding of all of this. There are just not enough hours in the day to learn it all.

[1] Perplexity of fixed-length models
[2] https://github.com/openai/openai-cookbook/blob/main/text_editing_examples.md#translation

5 Likes
6

Thanks for sharing a new way of creating a story. If I want to be more specific I also prompt “write a simple crime/science fiction/horror story” instead of simply writing “write a simple story”. Question & Answer section reinforces your understanding and this is how formal tests are designed to assess the language level of a foreign language learner. Good share.

2 Likes
7

Yes, you can expand with genre, and even ask for tenses or to use specific words.

Edit: Also worth mentioning, but if you do not completely like the result you get back, you can ask Chatty to tweak whatever part of the output, and it is usually pretty good at it.

2 Likes
8

GPT states that the entirety of Wikipedia is in the training data, which is of some help. Interestingly, if I open a new chat in Platt, Chatty always responds in auf Deutsch first. Then will only answer op Platt if I ask it to.

Chatty is seemingly unable to write in SONT or Nysassiske Skryvwyse from what I can tell. It claims it can, and there is some variation between what it outputs and SASS, but it looks far more off to me than SASS content it writes.

This is the issue when a language doesn’t have a written standard, or worse, has 3.

I am also quite curious about the inner workings and suspect it operates similar to whisper in that regard. I typically do not actually engage with Chatty in English (and in many cases our chats are probably a case study in multilingual code switching torment), but I wrote the prompt I showed here to give people an example of what they can do to get useful results.

4 Likes
9

I have to read this more carefully but I think we should create a space where share prompts that work for learning languages as we do.

I was trying to create a prompt with back and forward conversation for example. Or prompts for correcting mistakes that are more accurate and so on. Maybe Discord (that’s a bit dead) could be a place?

EDIT: I’ve just posted a message on Discord

4 Likes
10

So Low Saxon is basically Geraman-Dutch

1 Like
11

Be aware that I was once shown some texts about Nordic gods in Icelandic, created by Chat-GPT, that had quite a bit of errors, so don’t expect the program to get Faroese texts right, since it’s a much “smaller” language than even Icelandic.

3 Likes
12

The prompt wasn’t working for me, ChatGPT kept giving me enumerated sentences even in the story, I changed it this way.
I can also see you have English translations but they are not specified in your prompt, probably that was coming from previous prompts on the same chat.

What kind of command this #. should give?

Write a simple story in French (level A2) that is 10 sentences long.
This story should be told twice. Once from the perspective of the main character, and once from a third party perspective.

Now, create a question and answer section of 8 questions.
Enumerate each question like this:
Sentence from or about the story.
A question that reformulates the sentence from or about the story.
The answer to the question.

2 Likes
13

In my case the prompt has always worked from a new chat, but if something goes amiss you can tell ChatGPT what to do differently and it usually fixes issues very easily. What did Chatty return to that prompt?

1 Like
14

The first 2 parts of the stories were also enumerated like the questions. So I eliminated the # and it went ok like in your examples (without the English translation because they are not in the prompt).

1 Like
15

With the #. part, it solved an issue I was having where Chatty just returned 1. over and over again.

  1. Statement
    Question
    Answer

  2. Statment
    Question
    Answer

Etc.

1 Like
16

There should be a reader beware disclaimer on anything ChatGPT outputs at this time, or any AI models for that matter. The use cases I really see it are where ChatGPT can get a high percent accurate thing accomplished in a fraction of the time it will take a person.

In the case of simple short story writing, the % it can do on its own will be mostly dependent on the amount and quality of training data it gets. In the case of Faroese, probably less than Icelandic. The content it outputs should be checked and corrected by a native speaker or used by a learner that is going to be able to manage the inaccuracies and hallucinations produced by these models.

1 Like
17

As in the dialect continuum it sits geographically between Dutch and German it will look the part, especially in SASS orthography, but I would say they are more like first cousins to Dutch and German. If I were to extend the metaphor, Low Saxon is siblings with Frisian and English, but English got adopted by some Roman senator, kidnapped by some viking, before finally being adopted again by some French lord. Now English is a little… weird.

Low Saxon is also very interesting to me due to its impact on Swedish and Danish (and by extension Norwegian).

1 Like
18

Amazing, thanks noxialisrex. I’ve used this format to generate “mini stories” for Japanese. It works great to add also a topic to the prompt. Example for “fitness” below:

はい、以下にフィットネスをテーマにした、主人公の視点と第三者の視点の両方から語られる10文のシンプルストーリーを用意しました。

主人公の視点：
私は今日もフィットネスジムに行きました。トレッドミルでランニングをして、筋トレもしました。汗をかきながら、自分自身にチャレンジするのは気持ちがいいです。終わった後、自分が成し遂げたことに誇りを感じました。

第三者の視点：
彼は今日もフィットネスジムに行きました。トレッドミルでランニングをして、筋トレもしました。汗をかきながら、自分自身にチャレンジするのは気持ちがいいです。終わった後、彼は成し遂げたことに誇りを感じました。
質問と回答：

  1. 主人公は何をするためにフィットネスジムに行きましたか？ トレッドミルでランニングをしたり、筋トレをするために行きました。
  2. 主人公がフィットネスをすることによって得ることができるものは何ですか？ 自分自身にチャレンジすることや、成し遂げたことに誇りを感じることができます。
  3. 主人公はフィットネスジムで何をしましたか？ トレッドミルでランニングをしたり、筋トレをしました。
  4. 主人公はトレッドミルで何をしましたか？ ランニングをしました。
  5. 主人公は筋トレで何をしましたか？ 筋トレをしました。
  6. 主人公はフィットネスをすることによってどのような気持ちを得ることができますか？ 自分自身にチャレンジすることや、成し遂げたことに誇りを感じることができます。
  7. 主人公が成し遂げたことに対して感じた感情は何ですか？ 誇りを感じました。
  8. フィットネスをすることのメリットは何ですか？ フィットネスをすることによって健康を維持することができたり、自分自身にチャレンジすることや成し遂げたことに誇りを感じることができます。
1 Like
19

Worked great in french and danish. For danish i made it a scary mini story

20

I’ve actually had a conversation with it a few months ago about a similar thing, but here’s what what it said specifically about low Saxon and some of your questions (mostly unedited):

Low Saxon

Handling less common languages

My ability to use Low Saxon is based on my training on a large corpus of text in that language, which enables me to generate text that is grammatically correct and contextually appropriate. The training data is provided by humans with expertise in the language and culture, and my learning is based on statistical patterns in the text.

In general, my understanding of a language is not limited to specific topics, as I am designed to be a general-purpose language model. However, my ability to generate text in a specific language depends on the quality and quantity of training data that is available for that language.

To clarify, when generating text in Low Saxon or any other language, I use the knowledge and patterns that I have learned from the training data for that language. However, my understanding of the topic being discussed is not limited to that language, as I have access to a vast amount of information in multiple languages. This allows me to generate text that is accurate and relevant to the topic being discussed, regardless of the language in which it is being generated.

Translating internally

Yes, as an AI language model, I can internally translate information from one language to another, as I have been trained on a vast amount of data in multiple languages. This allows me to understand concepts and information in one language and generate text in another language that accurately conveys the same meaning.

When generating text in a specific language, I use my understanding of the concepts and information that are relevant to the topic being discussed, regardless of the language in which that information is presented. This allows me to generate text that is accurate and relevant, even when working with multilingual data.

However, it’s important to note that the accuracy of the translation can vary depending on the quality and quantity of training data that is available for the languages being used. Additionally, some concepts and ideas may not have an exact equivalent in other languages, which can make translation more challenging.

You should ask it more, especially the more specific questions :wink: And I’ve used the version 3.5 but remember that Bing has free ChatGPT 4 access in its search engine so you might get better answers there.

1 Like