Self-deprecating Humor Detection: A Machine Learning Approach

Abstract

Humor is one of the figurative language categories, and it is mainly used in human communication to express emotions and sentiments. Due to the complex structure of humorous texts, automatic humor detection is a challenging task. The detection becomes more challenging when we consider self-deprecating humor, which is a special category of humor in which users generally criticize and put themselves down. Interestingly, in recent years self-deprecating humor has been seen as a new business marketing trend, such as brand endorsement, advertisement, and content marketing. In this paper, we propose a novel self-deprecating humor detection approach using machine learning technique with an aim to enhance self-deprecating humor based marketing strategies. We have identified 16 new features related to three different feature categories – self-deprecating pattern, exaggeration, and word-embedding, and considered 11 humor-centric features from baseline works, and trained random forest classifier for detecting self-deprecating humor in Twitter. The proposed approach is evaluated over Twitter and two baseline datasets, and it performs significantly better in terms of standard information retrieval metrics.