When humour hurts: linguistic features to foster explainability

Lucia I. Merlo, Berta Chulvi, Reyner Ortega-Bueno, Paolo Rosso


The main objective of this research is to use different features for the textual representation of humorous texts and detect which are the characteristics that distinguish non-offensive jokes from the highly offensive ones. For this purpose, we use the data from the HaHackaton task in which jokes are annotated according to their degree of offensiveness. A new classification task is created by using two subsets of the jokes: the non-offensive ones and the highly offensive ones.The features with statistically significant differences in the two groups are used. By applying an ablation test, the most relevant features are used for a second classification task, showing that it is possible to obtain the same results with fewer computational resources.

Texto completo: