A Full Pipeline for Context-Aware Pun Generation

This paper proposes a full pipeline for Pun Generation, including the creation of the pun and alternative word pairs from a given context. Our experiments were carried out in Portuguese.

How to install

This project uses Python 3.9.

If you are using uv, run the following command:

uv init --python=3.9
uv sync

If you are using pip:

pip install -r requirements.txt

Afterward, remember to install the NLTK WordNet corpus:

import nltk
nltk.download('omw-1.4')

How to run

The pipeline's implementation (keyword extraction and expansion, homograph and homophone identification) is in the src folder. Our experiments, including NLG methods, results analysis, and other toy projects are in the scripts folder. The most important scripts are:

• scripts/generation/generate_ollama_jokes.py: Generate jokes using Ollama LLMs
• scripts/generation/generate_t5_jokes.py: Fine-tune and generate jokes using T5

All scripts require the data to be in the data folder in the data/processed_headlines.jsonl file, which already includes all pun and alternative signs created, in JSONL format. To create this file, you can use the scripts/preprocessing/preprocess_headlines.py script.

Generation results are saved in the results/generation/ folder, separated by the generation method.

Evaluation interface

The evaluation interface implementation is in the evaluation_interface folder, which requires streamlit to run. All evaluation results are in the results/evaluation/ folder, separated by evaluator. More information on how to configure the evaluation interface can be found in its own README file.

How to cite

@inproceedings{InacioMarcioLimaGoncaloOliveira2024,
    title     = {A {Full Pipeline} for {Context-Aware Pun Generation}},
    booktitle = {Proceedings of the 16th {{International Conference}} on {{Computational Creativity}} ({{ICCC}}'25)},
    author    = {In{\'a}cio, Marcio Lima and Gon{\c c}alo Oliveira, Hugo},
    year      = {2025},
    month     = jun,
    publisher = {Association for Computational Creativity (ACC)},
    address   = {Campinas}
}

Inácio, M. L., & Gonçalo Oliveira, H. (2025, June). A Full Pipeline for Context-Aware Pun Generation. Proceedings of the 16th International Conference on Computational Creativity (ICCC’25). 16th International Conference on Computational Creativity (ICCC’25), Campinas.