Clark Kent at SemEval-2023 Task 5: SVMs, Transformers, and Pixels for Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1204–1212 July 13-14, 2023 ©2023 Association for Computational Linguistics Clark Kent at SemEval-2023 Task 5: SVMs, Transformers, and Pixels for Clickbait Spoiling Dragos-Stefan Mihalcea Sergiu Nisioi dragos.mihalcea@s.unibuc.ro sergiu.nisioi@unibuc.ro Human Language Technologies Research Center Faculty of Mathematics and Computer Science University of Bucharest Abstract In this paper we present an analysis of our ap- proaches for the 2023 SemEval-2023 Clickbait Challenge. We only participated in the sub-task aiming at identifying different clikcbait spoil- ing types comparing several machine learning and deep learning approaches. Our analysis confirms previous results (Hagen et al., 2022) on this task and show that automatic methods are able to reach approximately 70% accuracy at predicting what type of additional content is needed to mitigate sensationalistic posts on social media. Furthermore, we provide a quali- tative analysis of the results, showing that the models may do better in practice than the met- ric indicates since the evaluation does not de- pend only on the predictor, but also on the ty- pology we choose to define clickbait spoiling. 1 Introduction There are many posts in social media that are in- tended to allure the readers into visiting a specific web page. The phenomenon is commonly known as clickbait and it is usually achieved through sen- sationalistic formulations that strike curiosity into the readers mind. Studies show that humans get dopamine rewards during information seeking processes driven by cu- riosity or pleasure of anticipation (Polman et al., 2022; Kobayashi and Hsu, 2019). Clickbait titles are linguistic exploits that activate these mecha- nisms and therefore users tend to reduce the per- ceived information gap (Loewenstein, 1994) by clicking or accessing the advertisements. While these type of linguistic exploits may be used to steer people into making healthier choices such as avoiding unhealthy foods, making more exercises etc. (Polman et al., 2022), most often, the information overload of clickbait posts from social media poses serios mental health risks (Hwang et al., 2021) and additional risks of propagating misleading information for profit or by the far-right for questionable political purposes (Ecker et al., 2014). Clickbait spoiling implies rephrasing the origi- nal title or the introduction of additional content in the original post with the purpose of limiting the cu- riosity of the readers. A techno-solutionist pipeline to spoilling for social media can be described by the following steps: 1. identify potential clickbait posts on social me- dia using text classification models 2. use a bot to access the clickbait websites re- trieved at the previous step and extract the corresponding content (linked document) 3. identify what type of information is required to reduce or limit the degree of curiosity of the original social media post 4. use the type of information and the linked doc- ument identified at the previous steps to gen- erate additional content acting as a spoiler The first step, for detecting whether a post is clickbait or not, has been greatly explored in re- cent years including …

Content truncated to 500 words