Buhscitu at SemEval-2020 Task 7: Assessing Humour in Edited News Headlines Using Hand-Crafted Features and Online Knowledge Bases

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 824–832 Barcelona, Spain (Online), December 12, 2020. 824 Buhscitu at SemEval-2020 Task 7: Assessing Humour in Edited News Headlines using Hand-Crafted Features and Online Knowledge Bases Kristian Nørgaard Jensen IT University of Copenhagen krnj@itu.dk Nicolaj Filrup Rasmussen IT University of Copenhagen nicr@itu.dk Thai Wang IT University of Copenhagen twan@itu.dk Marco Placenti IT University of Copenhagen mapl@itu.dk Barbara Plank IT University of Copenhagen bapl@itu.dk Abstract This paper describes our system to assess humour intensity in edited news headlines as part of a participation in the 7th task of SemEval-2020 on “Humor, Emphasis and Sentiment”. Various factors need to be accounted for in order to assess the funniness of an edited headline. We propose an architecture that uses hand-crafted features, knowledge bases and a language model to understand humour, and combines them in a regression model. Our system outperforms two baselines. In general, automatic humour assessment remains a difficult task. 1 Introduction Humour aims at generating amusement and laughter and can for this reason be considered one of the features enabling the creation of relationships in the interactions between humans. Understanding humour requires factual knowledge, context comprehension and—arguably—intelligence. Multiple factors play a role in the definition of humour, such as geographical location, culture, level of education and many others. This obviously makes the task of humour detection very hard for machines and artificially intelligent systems. In recent years, researchers operating in the field of computational linguistics have started to look into the topic, and a lot of progress has been made since the seminal paper by Mihalcea and Strapparava (2005). However, the quality of data sets leaves many questions unanswered, mainly because they are made of single punchlines or because sentences are divided into binary categories. Hossain et al. (2019) makes a remarkable effort on creating a data set of edited headlines where each headline is assigned a score representing the intensity of humour of that headline. This innovative data set enables researchers to conduct studies on a more granular level and may unlock novel techniques to get closer to a more efficient and successful computational model of humour. In this paper we propose an architecture that accounts for multiple factors that we believe play an important role in detecting the intensity of humour in a headline. In order to analyse the sentences, we include hand-crafted features extracted from the sentence itself and enable the system to look for the meaning of unknown objects using NELL (Never-Ending Language Learning) (Mitchell et al., 2015). This paper is a description of a system providing a solution to the SemEval2020 Task 7 (Hossain et al., 2020), which ranked 22nd out of 49 teams. The task is comprised of two sub-tasks. The first task is a regression task aimed at predicting the humour intensity of an edited headline. The second one is a classification task in which it is required to select the funnier headline out of the two provided. Our main focus was on the …

Content truncated to 500 words