ML Mob at SemEval-2023 Task 5: “Breaking News: Our Semi-Supervised and Multi-Task Learning Approach Spoils Clickbait”

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1818–1823 July 13-14, 2023 ©2023 Association for Computational Linguistics ML Mob at SemEval-2023 Task 5: "Breaking News: Our Semi-Supervised and Multi-Task Learning Approach Spoils Clickbait" Hannah Sterz∗ Leonard Bongard∗ Tobias Werner∗ Clifton A. Poth Martin B. Hentschel Technical University of Darmstadt {hannah.sterz, leonard.bongard, tobias.werner}@stud.tu-darmstadt.de Abstract Online articles using striking headlines that promise intriguing information are often used to attract readers. Most of the time, the infor- mation provided in the text is disappointing to the reader after the headline promised exciting news. As part of the SemEval-2023 challenge, we propose a system to generate a spoiler for these headlines. The spoiler provides the in- formation promised by the headline and elimi- nates the need to read the full article. We con- sider Multi-Task Learning and generating more data using a distillation approach in our system. With this, we achieve an F1 score up to 51.48% on extracting the spoiler from the articles. 1 Introduction The modern web is flooded with clickbait head- lines and articles - text snippets that try to lure the reader into clicking a link (Potthast et al., 2016) - so readers often fall prey to misleading information. Clickbait spoiling is a pressing issue that refers to the phenomenon of clickbait headlines that are used to gain attention and clicks to an article. The answer to the question or statement raised in the headline is typically a small part of the article. Clickbait spoiling, the process of getting the an- swer (the spoiler) to the raised problem in the ar- ticle, provides this small part of the article which enables readers to get the relevant information with- out having to go through the entire article. This work is part of the SemEval 2023 challenge where our team chose to focus on the topic of clickbait spoiling (Fröbe et al., 2023a), as it is a widespread problem that affects millions of internet users every day. The first of two tasks is about spoiler type classifi- cation, where an input text is given and the model classifies the spoiler as either "phrase", "passage" or "multi". A phrase spoiler denotes an article, whose headline can be answered with a single word ∗* equal contribution or short phrase. For "passage" spoilers the answer is a few sentences long. Answers with multiple spans are classified as the "multi" spoiler type. The second task focuses on spoiler generation. A clickbait post and a linked document are used as input to generate a spoiler (Hagen et al., 2022a). Within our work, we adopted the proposed base- lines by Hagen et al. (2022a). Specifically, we ap- plied Multi-Task Learning and extended the train- ing dataset by incorporating additional data. 2 Background This section introduces relevant background for the distillation approach and multi-task learning. 2.1 Clickbait Clickbait posts attempt to attract users with a con- troversial or striking title (Zannettou et al., 2022). The motivation behind this is to gain traffic on the site …

Content truncated to 500 words