THInC: A Theory-Driven Framework for Computational Humor Detection

TL;DR

This paper marks a pioneering effort in creating a humor detection framework that is informed by diverse humor theories and offers a foundation for future advancements in theory-driven humor classification, called THInC (Theory-driven Humor Interpretation and Classification).

Abstract

Humor is a fundamental aspect of human communication and cognition, as it plays a crucial role in social engagement. Although theories about humor have evolved over centuries, there is still no agreement on a single, comprehensive humor theory. Likewise, computationally recognizing humor remains a significant challenge despite recent advances in large language models. Moreover, most computational approaches to detecting humor are not based on existing humor theories. This paper contributes to bridging this long-standing gap between humor theory research and computational humor detection by creating an interpretable framework for humor classification, grounded in multiple humor theories, called THInC (Theory-driven Humor Interpretation and Classification). THInC ensembles interpretable GA2M classifiers, each representing a different humor theory. We engineered a transparent flow to actively create proxy features that quantitatively reflect different aspects of theories. An implementation of this framework achieves an F1 score of 0.85. The associative interpretability of the framework enables analysis of proxy efficacy, alignment of joke features with theories, and identification of globally contributing features. This paper marks a pioneering effort in creating a humor detection framework that is informed by diverse humor theories and offers a foundation for future advancements in theory-driven humor classification. It also serves as a first step in automatically comparing humor theories in a quantitative manner.

Figures (6)

• Figure 1: Graphical representation of the architecture of the THInC framework. The architecture consists of three different modules: feature generation, GA$^2$M classifiers and classification results, and interpreting the learned feature application in a classifier on its efficacy of capturing its humor theory. The symbol indicates the possibility of repeating (consecutive) flows over the adjacent quantification.
• Figure 2: The feature function of 'maximal change in anger' in the incongruity theory classifier is shown in black, with the minimal and maximal logit values of the bagged trees in gray, serving as uncertainty intervals. One possible numerical representation of the hypothesis -- that an increase in anger change correlates with a higher likelihood of incongruity -- is presented in blue, across all possible values an instance can have. There is a high correspondence between the actual feature function and the hypothesized feature function. The logit contribution of this feature in test joke 194 (formatted in bold in Figure \ref{['fig:localexample_barplot']}), corresponding to a specific value for this feature, is marked by an orange dot.
• Figure 3: The feature function of 'slope of linear fit of optimism' in the relief theory classifier is shown in black, with the minimal and maximal logit values of the bagged trees in gray, serving as uncertainty intervals. One numerical representation of the hypothesis that the slope of linear fit should be positive for the optimism (as a proxy for relief) to increase, is presented in blue across all possible values. There is a low correspondence between the actual feature function and the hypothesized feature function.
• Figure 4: The logit contribution of the most contributing features in the incongruity theory classifier to test joke 194: If I was a vampire hunter, I'd kill the vampires by inviting them over to my house and serving garlic bread. No one can resist that stuff. Orange contributions are negative, whereas blue contributions are positive. The intercept is the logit of the prediction that the model will make when all the features take their average values. The feature 'maximal change in anger' is formatted in bold, referring to a specific evaluation of the feature function in Figure \ref{['fig:featurefunction']}.
• Figure 5: The time series representing the anger probabilities of test joke 194 (Figure \ref{['fig:localexample_barplot']}), using a subsequence-based approach and the TweetNLP emotion classifier camacho-collados-etal-2022-tweetnlp. The maximal change of two anger probabilities is 0.78.