Comparing Jokes with NLP: How Far Can Joke Vectors Take Us?

摘要

This paper describes the results of comparison of jokes based on neural network language models and the General Theory of Verbal Humor knowledge resources. The question to be explored is whether joke similarity can be inferred from joke text similarity. Our results indicate that while similarities in text can be detected by doc2vec and jokes with somewhat similar texts can be consistently clustered, there is very little, if any, correlation between doc2vec cosine similarity and KR-annotated joke similarity.