Diane Simmons at SemEval-2023 Task 5: Is it possible to make good clickbait spoilers using a Zero-Shot approach? Check it out!

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 477–481 July 13-14, 2023 ©2023 Association for Computational Linguistics Diane Simmons at SemEval-2023 Task 5: Is it possible to make good clickbait spoilers using a Zero-Shot approach? Check it out! Niels Krog and Manex Agirrezabal Centre for Language Technology (CST) Nordic Studies and Linguistics (NorS) University of Copenhagen niels.krog1@gmail.com, manex.aguirrezabal@hum.ku.dk Abstract In this paper, we present a possible solution to the SemEval23 shared task of generating spoil- ers for clickbait headlines. Using a Zero-Shot approach with two different Transformer archi- tectures, BLOOM and RoBERTa, we generate three different types of spoilers: phrase, pas- sage and multi. We found, RoBERTa pretrained for Question-Answering to perform better than BLOOM for causal language modelling, how- ever both architectures proved promising for future attempts at such tasks. 1 Introduction Most of the companies that want to advertise them- selves use the internet as means of contact to the customers. The reputation of such companies is a strong signal of information quality, but social information encoded in metrics like the number of likes or shares on a post act as an alternative source of credibility (Munger, 2020). In their work, Munger (2020) call this environment Clickbait Me- dia. This content, whether it is produced by mass media, large companies or single individuals, is more likely to be shared if it evokes high-arousal emotions (Berger and Milkman, 2012). In this context, a clickbait is some internet con- tent whose main purpose is to encourage users to follow or click on a link to a web page. In a study conducted by Molyneux and Coddington (2020), the authors observed that stories with click- bait headlines were consistently judged to be of lower quality and lower credibility, but these differ- ences lacked of statistical power in the majority of the cases. Clickbait creates excitement in us, with devices such as cataphora, a type of anaphora, where the speaker mentions a general expression that will be further specified later in the utterance. An exam- ple of this is the following: “If you want some, here’s some parmesan cheese”,1 where the utter- ance parmesan cheese comes at the end of the sen- tence. Other methods for attracting users include sensationalism, exaggeration, and so on. The number of clicks in a website is a com- mon metric of user engagement (Lehmann et al., 2012). In the information retrieval community it was shown that a user sees a document almost always after clicking it (Dupret and Piwowarski, 2008). Therefore, we can expect the number of clicks in a website to be associated with a higher success. These clicks are commonly associated with economical revenue, which in some cases can be the economical support of many websites or platforms.2 In this paper we present our approach for the shared task on clickbait spoiling (Fröbe et al., 2023a; Hagen et al., 2022).3 Following current trends in Natural Language Processing (NLP), we approached the problem of clickbait spoiling as a Zero-Shot Question Answering …

Content truncated to 500 words