Computer-generated Humour Based on GPT-2

Abstract

Humour generation has always been a huge challenge in the area of computational humour. In this paper, we explore how to generate jokes based on keywords by fine-tuning GPT-2 pre-training model based on keywords task and comparing them with the LSTM-based encoder-decoder model. We trained the model by the short jokes of Conan O'Brien and the puns dataset of Yang et al. with the help of Pos-Tagger. Then, we evaluate the humour by using human evaluation and the similarity with keywords and jokes by using automatic evaluation. In terms of the final average score, the performance of our model is better than the LSTM-based encoder-decoder model.