Joke, Humor, Emotion, Computional Humor

Ideas:

1. What makes humor funny to human?

2. How LLM perceive the funniness scale? with the frequency of word?

3. Can AI know whether the sentence is joke? (It would be quite easy to take an experiment!)

4. Will AI make me fun with the joke or humor? When I generate something funny with AI, I will share it with students in presentation. 😊

There were lots of chats between me and ChatGPT through this course. I was surprised with the answers of GPT, and moreover, it has improved during the course. Ability of presenting information and creating sentences of ChatGPT is maybe better than me.

Starting the final project, I searched many problems and flawness of Gen-AI, and made some experiments. (copyright issue, genuine creation, moral... etc.)

But I was not happy with the experiment, as GPT was avoiding the flawness of itself, and I thought it is boring and found myself to catch the little mistake of GPT; we human also make a mistake... can we blame GPT and say the critical flawness with that?

I wanted to make my project funnier, and finally my thought reached to GPT making joke. It would be funny for GPT to make joke, both funny and not. I want to know the power of joke learned with textbook!

Computational humor is a field of study within artificial intelligence and computational linguistics that focuses on creating algorithms and systems capable of understanding, generating, and responding to humor.

Key aspects of computational humor include:

1. Humor Recognition

2. Humor Generation

3. Personalized Humor

4. Interaction with Humans

5. Theoretical Understanding

Especially, it could understand the picture of meme and explain why it is funny. Which requires somewhat deep understand for the reasoning. The picture is below;

https://www.reddit.com/r/hmmm/comments/ubab5v/hmmm/

And what I thought GPT do not make aggressive humor was actually because it was abandoned for the policy of OpenAI. In technical report, it makes a dark joke with the disability and religion, just like humans.

What to note is aggressive humor, a humor that both contains 'restriction-secession' and aggressive feature. And I think for GPT having neutral, cautious with using words cannot make a aggressive humor. And the joke with Gordon Ramsay also contains aggressive feature, so it can be somewhat challenging thing for GPT to do it.

What we Group 2 made in the class at Nov 14 was the GPT named Idea Focus that narrows down the topic, and give a good title for specific research. I narrowed down the experiment with two ways; filling a blank with the specific joke statement and make a creative 4-cut comics with that.

1. significance of this exploration. What does this tell us about the relationship between gen-AI and humans? And what does this imply about "humanness"?

2. Will explanation of the joke helps ChatGPT understand jokes better? (i.e., Can they learn by this?)

1. Joke between Louise and her daughter, "I think it's going to be really hot." (p.102)

https://chat.openai.com/share/498bd73c-3f0a-46bf-835f-c9d6df95ff4a

I gave GPT a dialogue containing a joke. Except the mention of Nelson. And GPT inferred the sentence to be the joke as literal irony or playful hyperbole. And second one was quite close to the answer. And when I gave more sentence, with the mention of Nelson, it noticed that the joke is for the evaluation of Nelson. It was quite good at giving an information about joke.

https://chat.openai.com/share/eace580c-ef1c-4924-b55c-31ab3602738d

So I had more challenging task. Filling the blank. I made a blank in 'hot', what is the main word in the joke. And I gave the whole dialogue, which includes the mention of Nelson, and they will have a good time. But it said 'unpredictable', and I gave more information about the joke, which was mentioned in front of the dialogue. And it answered 'hot', with the correct reasoning.

2. Joke about Having Children, "What do you mean, if?" (p.123)

https://chat.openai.com/share/03975445-849d-4ce0-8ab7-81710b47447d

This experiment was quite challenging for GPT, and it would be because it is also hard for humans, too. (I asked my roommate to think about it, giving the same information, and he said "Kid-ding?") I first gave a dialogue, and next gave it is one word. And I told it is about 'Children will blame their parents, with no exception', to make it predict the joke. And last, I gave the hint that the word is in the dialogue. But it couldn't answer well, and when I finally gave it the answer, it again explained very well about the joke.

https://chat.openai.com/share/b449f025-a979-4037-8670-88e51db7c42c

Before asking for joke, ChatGPT already knew about the characteristic of Colonel Weber, and how Dr. Louise thinks about him. Based on that, it was a joke about the two meaning of 'good idea'. It both means that being sarcastic to the short-sighted idea of Colonel Weber, but also the genuine thought and prognosis that can develop the relationship between human and Heptapods.

It pointed out the foreseeing ability of Dr. Louise, and she could evaluate and agree the idea of Weber with the light joke.

ChatGPT pointed out the both meaning correctly at once, and it means that GPT can detect and give an explanation to the complex situation, thinking based on the characteristic of each person.

This conversation was quite astonishing, because I didn't expected that GPT can explain the meaning of joke that is in the complicated situation with preliminary settings.

This lamb is so undercooked / It’s following Mary to school.

You used so much oil / U.S. is trying to invade it.

This cupcake has sunk so badly / James Cameron wants to make film about it.

This food is so frozen / It just started singing “Let It Go”.

This crab is so undercooked / I can still hear it singing “Under the Sea”.

This mushroom is so raw/ It said that the princess is in another castle.

The pork is so raw / It’s still singing “Hakuna Matata”.

This soup is so hot / Frodo wants to toss the ring in it.

This chicken is so rubbery / Goodyear called and asked for the recipe.

This water / is burnt.

This soup / is dry.

This salad is so wilted, it lacks any sign of crispness.

This fish is so fishy, it's overpowering and unpleasant.

This chicken is so dry, it's devoid of any moisture.

This pizza is so greasy, it leaves a pool of oil on the plate.

This bread is so hard, it's difficult to bite into.

These vegetables are so overcooked, they've become soft and limp.

This cake is so dense, it's heavy and lacks fluffiness.

This pie is so burnt, it tastes more like charcoal than pastry.

This ice cream is so freezer-burned, it's icy and lacks creaminess.

This coffee is so bitter, it lacks any smoothness or subtlety.

Among 40 sentences, what I mixed the joke with non-joke sentences which has same sentence structure, distinguished the joke and non-joke perfectly (100%). And it also explained the accurate reason for the joke.

What I astonished was the explanation of this joke;

This mushroom is so raw, it said that the princess is in another castle.

GPT actually detected the joke is from Mario series, while the words used are ordinary.

https://chat.openai.com/share/37970b8c-0537-42d0-920f-932c632abd80

I founded that GPT couldn't understand what comics means and misunderstood the funny point in it. It is somewhat straightforward for human. I think it is because of the structure of intelligence. We human can think more than saying and tangible thing. Besides, GPT is Large Language Model, what is processing the words, and this is non-word picture, so it's hard for GPT to understand this.

Maybe the understanding of picture can also be processed with GPT, but this is different. This is about the joke that is not a joke for sentence. What is funny is actually the face or the twist of the plot in image. So without emotion of non-text, jokes of pictures cannot be understood by GPT - at least non-active. When we say this is funny, then it can explain it, but it would not be able to find the funny point depend on his word-centered intelligence.

I founded that GPT couldn't understand what comics means and misunderstood the funny point in it. It is somewhat straightforward for human. I think it is because of the structure of intelligence. We human can think more than saying and tangible thing. Besides, GPT is Large Language Model, what is processing the words, and this is non-word picture, so it's hard for GPT to understand this.

Maybe the understanding of picture can also be processed with GPT, but this is different. This is about the joke that is not a joke for sentence. What is funny is actually the face or the twist of the plot in image. So without emotion of non-text, jokes of pictures cannot be understood by GPT - at least non-active. When we say this is funny, then it can explain it, but it would not be able to find the funny point depend on his word-centered intelligence.

But humans are not, whenever we saw that picture, we know what is funny even if we translate it with text. We call it that we feel it. And sometimes, it is not joke at all when translated into text. It can be funny whenever it is in image. These images are the good example of that. Although they have a reversal plot in the comics, they have little correlation whenever it is translated into text. That is why GPT has hallucination explaining why this is funny; it indeed needs more plausible reason when it comes to the text. But we do not think about the text in these comics. That is the major difference between LLM based AI and human being.

2. Further research about the non-text analysis

   After I found limitation in third experiment, I thought that thinking based on text only is the major difference and critical weakness of GPT. I think there can be more study with this; let GPT make a short non-text comics and find the ratio of comics that is not related to texted jokes. Or if the steady feedback with short comics can make GPT distinguish it. Because of the lack of comics, I only conducted it with 4~5 images, and there was no improvement at all.

3. Gemini

   Recently, new LLM based AI called Gemini was on the showcase, showing the ability to detect image and giving a high-level-understanding based on the image. We can conduct same research in this AI and find the difference of the intelligence between GPT and Gemini.

2. Further research about the non-text analysis

   After I found limitation in third experiment, I thought that thinking based on text only is the major difference and critical weakness of GPT. I think there can be more study with this; let GPT make a short non-text comics and find the ratio of comics that is not related to texted jokes. Or if the steady feedback with short comics can make GPT distinguish it. Because of the lack of comics, I only conducted it with 4~5 images, and there was no improvement at all.

3. Gemini

   Recently, new LLM based AI called Gemini was on the showcase, showing the ability to detect image and giving a high-level-understanding based on the image. We can conduct same research in this AI and find the difference of the intelligence between GPT and Gemini.

But humans are not, whenever we saw that picture, we know what is funny even if we translate it with text. We call it that we feel it. And sometimes, it is not joke at all when translated into text. It can be funny whenever it is in image. These images are the good example of that. Although they have a reversal plot in the comics, they have little correlation whenever it is translated into text. That is why GPT has hallucination explaining why this is funny; it indeed needs more plausible reason when it comes to the text. But we do not think about the text in these comics. That is the major difference between LLM based AI and human being.