Computational Experiments on Detecting Meaning shift in Jokes

The paper describes an experimental approach to detect the meaning shift, one of the most fundamental characteristics of humor, which is studied by many scientists in different interdisciplinary theoretical methodologies. We measured cosine similarity between setups and punchlines and explained these results through the set of objective criteria such as cosine results limitations, punchline length, three groups of words in joke’s parts. We also decided to investigate the originality of obtained distribution of cosine similarity to the same distributions calculated for fiction texts. The results demonstrate crucial differences in distributions for all the verification texts. We described an automatic approach for extracting meaning concepts of setup and punchline aided by word embeddings of its top three semantically closest words. We also provide a comparative analysis of distribution of cosine similarities among jokes. The proposed approach allows obtaining embeddings for setup and punchline made on the top of the closest words in semantic space.