nancy-hicks-gribble at SemEval-2023 Task 5: Classifying and generating clickbait spoilers with RoBERTa

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1712–1717 July 13-14, 2023 ©2023 Association for Computational Linguistics nancy-hicks-gribble at SemEval-2023 Task 5: Classifying and generating clickbait spoilers with RoBERTa Jüri Keller TH Köln Claudiusstr. 1 50678 Köln jueri.keller@smail.th-koeln.de Nicolas Rehbach TH Köln Claudiusstr. 1 50678 Köln nicolas_alexander.rehbach@smail.th-koeln.de Ibrahim Zafar TH Köln Claudiusstr. 1 50678 Köln ibrahim.zafar@smail.th-koeln.de Abstract Clickbait spoiling and spoiler type classifica- tion in the setting of the SemEval2023 shared task five was used to explore transformer based text classification in comparison to conven- tional, shallow learned classifying models. Ad- ditionally, an initial model for spoiler creation was explored. The task was to classify or create spoilers for clickbait social media posts. The classification task was addressed by com- paring different classifiers trained on hand crafted features to pre-trained and fine-tuned RoBERTa transformer models. The spoiler gen- eration task was formulated as a question an- swering task, using the clickbait posts as ques- tions and the articles as foundation to retrieve the answer from. The results show that even of the shelve trans- former models outperform shallow learned models in the classification task. The spoiler generation task is more complex and needs an advanced system. 1 Introduction Clickbait describes the presentation of content on the internet in such a manner, that it encourages the viewer to click on a link1. To increase the likelihood of a click, lurid headlines or mislead- ing descriptions are oftentimes used. Clickbait spoiling oppositely tries to counteract by precisely summarising the most important information of a given article or document. Task five of Se- mEval2023’s Discourse and Argumentation section revolves around creating Clickbait spoilers based on clickbait posts on social media, organized and evaluated through (Fröbe et al., 2023b,a). Further 1Oxford Learner’s Dictionaries information about clickbait spoiling and the under- lying data of the challenge is presented in (Hagen et al., 2022). The Clickbait Challenge consisted of two tasks. The first is spoiler type classification and the sec- ond is spoiler creation. Based on a social media post and the given article the correct spoiler type shall be classified for the first task. For task two, the correct spoiler shall be generated based on the information provided such as the advertised article. Three systems based on the RoBERTa (Liu et al., 2019) transformer model were submitted for the first task: A multiclass classifier, a system based on three binary classifiers and a multiclass clas- sifier that was further pre-trained. Additionally, these systems were compared to different shallow learned classifiers beforehand. As for task 2, a sys- tem based on an off the shelve transformer model and spoiler type dependent post processing is pro- posed for the spoiler creation. The submitted systems are made publicly avail- able on GitHub2. 2 Background The dataset for the shared task is provided through the organizers website and contains 3200 rows for training and 800 for validation. 1000 additional spoilers were held back as testing data for the sub- mission. Out of 14 existing columns, …

Content truncated to 500 words