Authors: Yingchen Ma, Nemath Ahmed, Rajvi Parekh
In this project, we used fine-tuning and few-shot prompting techniques to improve the ability of large language models (LLMs) to understand humor. Focusing on puns (a specific category of humor), we conducted experiments to investigate if these techniques allow the LLM to generate explanations for these puns that closely resemble human-produced explanations and effectively communicate their humor. Subsequently, we performed human annotation to evaluate the clarity, completeness, relevance, and insightfulness of the LLM-produced explanations. Through both quantitative and qualitative evaluations, we demonstrate that both fine-tuning and few-shot prompting techniques improve a Llama-2-7b model's ability to generate high-quality explanations for puns.
The full project report, alongside the code and data, can be found in this repository.
This was a course final project for CS 8803 LLM (Large Language Models).
annotations/: human annotations for the quality of the LLM-produced explanations.data/: folder for input data.experiments/: notebooks containing all code.results/: files containing results.