85% Accuracy: AI Now Cracks Jokes Like a Human

Researchers from the University of Antwerp, Leuven, and Montreal have developed the THInC (Theory-driven humor Interpretation and Classification) model, which uses Generalized Additive Models (GAMs) and pre-trained language transformers to detect humor with remarkable accuracy. The model leverages three major humor theories: Incongruity, Superiority, and Relief. By converting text into time series and analyzing sentiment changes, the model identifies humor through numerical proxy features. The combined model achieved an F1 score of 85%, with individual models scoring between 79 to 81. This approach not only detects humor but also provides interpretability, showing how different emotions contribute to humor classification. Despite its success, the model’s proxy features could be refined to better handle the inherent noise in text.

Source: towardsdatascience.com