SemEval-2023 Task 5: Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2275–2286 July 13-14, 2023 ©2023 Association for Computational Linguistics SemEval-2023 Task 5: Clickbait Spoiling Maik Fröbe Friedrich-Schiller-Universität Jena Tim Gollub Bauhaus-Universität Weimar Benno Stein Bauhaus-Universität Weimar Matthias Hagen Friedrich-Schiller-Universität Jena Martin Potthast Leipzig University and ScaDS.AI Abstract In this overview paper, we report on the second PAN Clickbait Challenge hosted as Task 5 at SemEval 2023. The challenge’s focus is to bet- ter support social media users by automatically generating short spoilers that close the curiosity gap induced by a clickbait post. We organized two subtasks: (1) spoiler type classification to assess what kind of spoiler a clickbait post war- rants (e.g., a phrase), and (2) spoiler generation to generate an actual spoiler for a clickbait post. 1 Introduction Clickbait posts link to web pages and advertise their content by arousing curiosity instead of pro- viding informative headlines or summaries. The PAN Clickbait Challenge at SemEval 2023 aims to develop techniques for automatic clickbait spoil- ing: generating a short text that can close the cu- riosity gap as exemplified in Figure 1. Despite manual clickbait spoiling being popular on social networks,1 the problem had not received much at- tention by the research community. We thus or- ganized a respective challenge based on our pre- viously published dataset and baseline approaches (Hagen et al., 2022). With approaches submitted by 30 teams, the challenge had a very good participa- tion. To improve reproducibility and replicability, we asked the teams to submit their software via the TIRA platform (Fröbe et al., 2023)—already used in several shared task with an archive of more than 500 research prototypes. Still, for our Click- bait Challenge, we also allowed result submissions, but a majority of the teams (19 of the 30) actually submitted software in the form of Docker images. The task of clickbait spoiling, as exemplified in Figure 1, specifically aims at completing the “life cycle” of clickbait posts. The creation (Xu et al., 2019) and detection (Potthast et al., 2016, 2018) of clickbait posts are already operationalized, so we 1Saved You a Click has 1.9 million members on Reddit. New York Post @nypost Just how safe are NYC's water fountains? nyp.st/2yHSGnr “The Post independently tested eight water fountains in New York City’s most frequented parks, and found that all met or exceeded the state’s guidelines for water quality.” Lifehacker @lifehacker How to keep your workout clothes from stinking: lifehac.kr/57YOuEZ “washing [them]” Spoiler Clickbait tweet CNBC @CNBC A Harvard nutritionist and brain expert says she avoids these 5 foods that "weaken memory and focus." (via @CNBCMakeIt) cnb.cx/2TG6zeX “1. Added sugar” [...] “2. Fried foods” [...] “3. High-glycemic-load carbohydrates” [...] “4. Alcohol” [...] “5. Nitrates” [...] Figure 1: Clickbait tweets and spoilers from the linked web pages (phrase, passage, and multipart spoiler). now aim to cover the “end” by automating spoiling. If sufficiently many people adopt spoiling tools, a clear message is send to publishers and social media platforms. Still, spoiling clickbait instead of removing …

Content truncated to 500 words