Commonsense-Aware Attentive Modeling for Humor Recognition.

Laughter has positive effects on health. Humor is an important component of daily communication and usually causes laughter. Hence, we can expect that the infusion of humor in a dialogue system will improve the user’s physical and mental satisfaction. Since even humans have difficulty comprehending humor, appropriate knowledge is essential for humor understanding. In this paper, commonsense-aware modules are extrapolated to P re-trained L anguage M odel s (PLMs) to provide external knowledge. We specifically extract keywords from a text and use COMET to obtain embeddings that represent the commonsense associated with the keywords. We attempt to detect humor that is not detectable by PLM alone. Our approach enables the model to access commonsense knowledge. Compared to the baseline, the number of humor detections increases, and recall is improved without a significant decrease in precision. Our best model significantly improves recall by 4.4% for a 0.4% reduction in precision in the HaHackathon dataset and by 20.3% for an 8.4% reduction in precision in the Humicroedit dataset compared to the baseline. We also observe the changes in prediction and processing speed so as to analyze the characteristics of the proposed method and the issues for its social implementation.