Walter Burns at SemEval-2023 Task 5: NLP-CIMAT - Leveraging Model Ensembles for Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 693–699 July 13-14, 2023 ©2023 Association for Computational Linguistics Walter Burns at SemEval-2023 Task 5: NLP-CIMAT - Leveraging Model Ensembles for Clickbait Spoiling Emilio Villa-Cueva1 Daniel Vallejo-Aldana1 Fernando Sánchez-Vega1 Adrián Pastor López-Monroy1 1Department of Computer Science, Mathematics Research Center (CIMAT), Guanajuato, Mexico {emilio.villa,daniel.vallejo,fernando.sanchez,pastor.lopez}@cimat.mx Abstract This paper describes our participation in the Clickbait challenge at SemEval 2023. In this work, we address the Clickbait classification task using transformers models in an ensemble configuration. We tackle the Spoiler Genera- tion task using a two-level ensemble strategy of models trained for extractive QA, and selecting the best K candidates for multi-part spoilers. In the test partitions, our approaches obtained a classification accuracy of 0.716 for classifica- tion and a BLEU-4 score of 0.439 for spoiler generation. 1 Introduction Clickbait is the name given to content whose pri- mary purpose is to lure visitors to a web page. Often, this content is designed to appeal to users through sensationalism, aiming to affect their cu- riosity. Once the user takes the bait, very often contains a large text that has little to do with the promised content in the Clickbait, and there are only a few sentences that properly “answer” it. In their work, (D. Molina et al., 2021) describes that Clickbait texts rely on the cognitive phenomenon known as Information Gap, consisting of alerting the reader with some relevant information on the text but leaving a gap in the reader’s knowledge hence triggering its curiosity and the need to fill this gap. Although the purpose of clickbait is to appeal to the reader’s curiosity and emotions, this can have side effects as well. As stated in (Zhou, 2022), clickbaits usually exaggerate headlines to attract the attention of the audience ignoring completely the fact that news should follow such as objectivity, truth and fairness. This can mislead the readers to have a wrong judgment of news events. The Clickbait Challenge at SemEval 2023 (Fröbe et al., 2023a) is a shared task whose objective is to de- velop systems that, given a Clickbait and the con- tent of its linked page, can classify the required types of spoiler (task 1) and produce an adequate spoiler (task 2). There are examples of the use of automatic meth- ods for detecting Clickbaits (Potthast et al., 2016; Agrawal, 2016; Khater et al., 2018), in which they aim to develop a system that can identify Clickbait content and label it as such before it reaches the user. However, there is less research on systems for au- tomatically “answering” Clickbaits. In their work, Hagen et al. (2022) proposes “spoiling” the posts by treating the problem as either a passage-retrieval or a question-answering problem. They experimen- tally found there is a significant difference in favor of the question-answering approach. In the same work, the authors compile the “Webis Clickbait Spoiling Corpus 2022”, a dataset that consists of 5000 manually spoiled posts, and define three pos- sible categories of spoilers: multi-part, …

Content truncated to 500 words