Generating humorous statements for public speech with pre-trained language model and tension analyzer

Abstract

Proper humor, as used in public speeches or presentations, relieves tension between the audience and the presenter, creates a sense of closeness, and sustains the attention of the audience. Therefore, automatically suggesting proper humorous statements that fit the context of the presentation can be of help to the presenter. Generating context-based humor, however, has rarely been attempted in the existing humor studies. This thesis suggests a method of generating humorous statements consistent with a given context by transfer learning of the language model, which is pre-trained from a large amount of text. Also, to generate the ones suitable for a presentation, a tension analyzer is suggested to filter out those not helpful to the development of the presentation via a corpus created by annotating the tension development in TED Talks. Automated and human assessments show that our system can suggest humorous statements that are fun as well as consistent with the given context and help develop the presentation. A user interface and an application are introduced to suggest some humorous statements when a presentation script is typed in.