Chick Adams at SemEval-2023 Task 5: Using RoBERTa and DeBERTa to Extract Post and Document-based Features for Clickbait Spoiling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 624–628 July 13-14, 2023 ©2023 Association for Computational Linguistics Chick Adams at SemEval-2023 Task 5: Using RoBERTa and DeBERTa to extract post and document-based features for Clickbait Spoiling Ronghao Pan1, José Antonio García-Díaz1, Francisco García-Sánchez1, Rafael Valencia-García1 Facultad de Informática, Universidad de Murcia, Campus de Espinardo, 30100 Murcia, Spain {ronghao.pan@um.es, joseantonio.garcia8, frgarcia, valencia}@um.es Abstract In this manuscript, we describe the participa- tion of the UMUTeam in SemEval-2023 Task 5, namely, Clickbait Spoiling, a shared task on identifying spoiler type (i.e., a phrase or a passage) and generating short texts that sat- isfy curiosity induced by a clickbait post, i.e. generating spoilers for the clickbait post. Our participation in Task 1 is based on fine-tuning pre-trained models, which consists in taking a pre-trained model and tuning it to fit the spoiler classification task. Our system has obtained excellent results in Task 1: we outperformed all proposed baselines, being within the Top 10 for most measures. Foremost, we reached Top 3 in F1 score in the passage spoiler ranking. 1 Introduction Clickbait is a term used to describe social media posts that aim to attract users’ attention and get them to click on those posts or visit a website. This is achieved through catchy titles that often exagger- ate or distort the actual content of the post in order to generate traffic or advertising income (Hagen et al., 2022). Clickbait is considered inappropriate because its resolution is often ordinary or trivial, consisting of little more than a sentence, a short passage or a list of things that could easily have been included in the post. For this reason, the Clickbait Spoiling shared task (Fröbe et al., 2023a), proposed at SemEval 2023, consists of counteract- ing the negative effects of clickbait providing a spoiler or preview that reveals the main content of a post without requiring the user to click through to the website. The task is divided into two sub-tasks: 1. Spoiler type classification: The task is to classify the spoiler type that the clickbait post warrants. There are basically three types of spoilers: (1) phrase spoilers consisting of a single word or phrase from the linked docu- ment, (2) passage spoilers consisting of one or a few sentences of the linked document, and (3) multipart spoilers consisting of more than one non-consecutive phrases or passages of the linked document. 2. Spoiler generation: The task is to generate the spoiler for the clickbait post. One of the novel features of this shared task is the integration of the TIRA platform for the repro- ducibility of tasks (Fröbe et al., 2023b), preventing third parties who want to evaluate the state of the art of a task on other datasets from having to re- implement the participants’ software (Fröbe et al., 2023). Advances in deep learning have had a major im- pact on Natural Language Processing (NLP), signif- icantly improving the ability of computation to un- derstand and process human language. Moreover, …

Content truncated to 500 words