"Humor" is always being a hard topic since it's highly personal and hard to measure. It's really cool if I can generate a comic from a joke. This project is the final project for the course "Conginitive Computing 2018". The goal of this course is "beyond recognition", so I wish to deliver a project that fits this goal.
- Image often contains more information that text. The relation between text and image is one-to-many, and the context also affects choosing "correct" image representation.
- There is no dataset nor previous work for such task.
Given a joke x, output a comic y
I defined two types of joke comics to lower the difficulty:
- The joke is a question and a humorous answer
- The comic will be an image with the joke itself.
- The joke is a conversation (mainly two people included)
- The comic will be an 4-frame cartoons describing the joke.
- For type-1 joke, the main task is to "select" a proper image.
- Additionally, maybe we can use GAN to generate the image instead of selecting it.
- For type-2 joke, the challenge is to split the joke and meanwhile maintaining the relation between each frames
- Crawl PTT Joke board
- Try to "select" image by the text embedding
- Croud sourcing to evaluate the result (generate 4 result, choose the best fits)
- (Additional) GAN the image
- Joke dataset https://github.com/taivop/joke-dataset