Generating Japanese Puns via Paraphrasing Using a Language Model Fine-Tuned with Furigana-Annotated Corpus

Abstract

Puns, a form of wordplay, involve the creation of sentences by combining phonetically similar but semantically different words. Generating puns requires a deep understanding of word meanings and pronunciations. In this study, we develop a model that can transform Japanese sentences into puns without altering their original meanings by fine-tuning pre-trained language models to focus on Japanese phonetics. Leveraging corpora annotated with furigana and a pun database, we enhance the language model's ability to grasp Japanese phonetic nuances and its capacity to generate puns. Comparative experimental results reveal that our method, in contrast to models fine-tuned solely with a pun dataset without considerations for phonetic comprehension, achieves a 0.035-point increment in the BLEU score.