John Boy Walton at SemEval-2023 Task 5: An Ensemble Approach to Spoiler Classification and Retrieval for Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2100–2106 July 13-14, 2023 ©2023 Association for Computational Linguistics John Boy Walton at SemEval-2023 Task 5: An Ensemble Approach to Spoiler Classification and Retrieval for Clickbait Spoiling Maksim Shmalts Department of Linguistics, University of Tübingen maksim.shmalts@student.uni-tuebingen.de Abstract Clickbait spoiling is a task of generating or re- trieving a fairly short text with a purpose to satisfy curiosity of a content consumer with- out their addressing to the document linked to a clickbait post or headline. In this paper we introduce an ensemble approach to clickbait spoiling task at SemEval-2023. The tasks con- sists of spoiler classification and retrieval on Webis-Clickbait-22 dataset. We show that such an ensemble solution is quite successful at clas- sification, whereas it might perform poorly at retrieval with no additional features. In con- clusion we outline our thoughts on possible directions to improving the approach and shape a set of suggestions to the said features. 1 Introduction Clickbait is a term for a descriptive piece of in- formation (a headline, a snippet, a post in social media etc.) aiming to arouse users’ interest by summarizing the contents of a linked document in a deceptive and/or misleading manner. The sum- mary does not suffice to satisfy users’ curiosity, whereas the linked document is in a way adver- tized to do so, which encourages viewers to click the link (and thus cover the arisen curiosity gap). Using clickbait is frowned upon as the contents of the linked page does not usually correspond to the expectations of users and in some cases may be inappropriate, offensive or harmful. Hagen et al. (2022) address the issue with an idea of clickbait spoiling: 1) predicting the type of a spoiler that would satisfy users’ curiosity and thus withhold them from following the link; 2) retrieving the spoiler itself from the document. Upon developing a solution Hagen et al. (2022) compiled a dataset containing "5,000 “spoilable” clickbait posts" in English collected from Twitter, Reddit, and Face- book, that then were manually spoiled and classi- fied: The Webis Clickbait Spoiling Corpus 2022 (Webis-Clickbait-22). SemEval-2023 Task 5 sug- gests to introduce a model for clickbait spoiling and bring new insights to the existing approaches, basing on Webis-Clickbait-22 and (optionally) ad- ditional resourses (Fröbe et al., 2023). Our team presents an ensemble solution called WebSemble (Webis-Clickbait-22 Ensemble) mainly following the steps outlined in Hagen et al. (2022) except for an additional (and optional) sum- marization of the documents before the classifi- cation. Below we briefly list actions undertaken with our approach upon execution (more in section System Overview): 1. Data splits are transformed into SQuAD1.1 (Rajpurkar et al., 2016) format and preprocessed; 2. The contents of the linked web pages are sum- marized if summarization is used; summarization is conducted with a single model; 3. The clickbait posts are spoiled: 3.1. Types of spoilers are predicted from either summarized documents (if summarization is used) or their titles; whether summarization is used is one …

Content truncated to 500 words