HumorAAC at SemEval-2020 Task 7: Assessing the Funniness of Edited News Headlines through Regression and Trump Mentions

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1019–1025 Barcelona, Spain (Online), December 12, 2020. 1019 HumorAAC at SemEval-2020 Task 7: Assessing the Funniness of Edited News Headlines through regression and Trump mentions Anna-Katharina Dick Charlotte Weirich Alla Kutkina University T¨ubingen {anna-katharina.dick, charlotte.weirich, alla.kutkina}@student.uni-tuebingen.de Abstract In this paper we describe the HumorAAC system, our contribution to the Semeval-2020 Humor Assessment task. We essentially use three different features that are passed into a ridge regression to determine a funniness score for an edited news headline: statistical, count-based features, semantic features and contextual information. For deciding which one of two given edited headlines is funnier, we additionally use scoring information and logistic regression. Our work was mostly concentrated on investigating features, rather than improving prediction based on pre-trained language models. The resulting system is task-specific, lightweight and performs above the majority baseline. Our experiments indicate that features related to socio-cultural context, in our case mentions of Donald Trump, generally perform better than context-independent features like headline length. 1 Introduction Humor is an essential part of natural everyday human communication. A huge range of research was devoted to the nature of humor in various fields such as psychology, linguistics and philosophy resulting in various theories such as incongruity-resolution theory, non-cooperative theory, superiority theory, release theories, semantic script theory and general theory of verbal humor (Attardo, 2008). None of these theories could give a complete and comprehensive explanation of humor phenomena so far. For the field of computational linguistics humor is being quite a difficulty as well. Because of its complexity and inherent subjectivity, the development of automatic humor recognition and assessment poses a great challenge in Computational Linguistics, and therefore is a popular subject in various shared task competitions (Chiruzzo et al., 2019; Hossain et al., 2020) and independent research (Reyes et al., 2012; Mihalcea and Strapparava, 2005; Zhang and Liu, 2014; Yang et al., 2015). This challenge is related to the meaning extraction task in the domain of sentiment analysis since a great deal of naturally produced text contains humor in one form or another. The system described in this paper was created for the humor assessment task within the International Workshop on Semantic Evaluation SemEval-2020 (Hossain et al., 2020). This task is aiming to develop a system that automatically evaluates the humorousness of the edited headlines, made funnier with short one-word edits, and evaluated by human annotators. We believe, that this task can be an important step on a way to the development of better speech recognition systems and to the understanding of humor as a phenomenon in general. The presented system is based on ridge regression and neural networks. In essence we pass a set of count-based, semantic, and context-oriented features into the ridge regression algorithm. The main discovery of our research is the importance of task- and data-specific features on the performance of a humor-detection system. Due to the social, context-dependent nature of humor, the addition of two features pertaining to Donald Trump, originally added only for …

Content truncated to 500 words