Gallagher at SemEval-2023 Task 5: Tackling Clickbait with Seq2Seq Models

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1650–1655 July 13-14, 2023 ©2023 Association for Computational Linguistics Gallagher at SemEval-2023 Task 5: Tackling Clickbait with Seq2Seq Models Tugay Bilgis, Nimet Beyza Bozdag, Steven Bethard University of Arizona {tbilgis, nbbozdag, bethard}@arizona.edu Abstract This paper presents the systems and approaches of the Gallagher team for the SemEval-2023 Task 5: Clickbait Spoiling. We propose a method to classify the type of spoiler (phrase, passage, multi) and a question-answering method to generate spoilers that satisfy the cu- riosity caused by clickbait posts. We experi- ment with the state-of-the-art Seq2Seq model T5. To identify the spoiler types we used a fine- tuned T5 classifier (Subtask 1). A mixture of T5 and Flan-T5 was used to generate the spoil- ers for clickbait posts (Subtask 2). Our system officially ranks first in generating phrase type spoilers in Subtask 2, and achieves the high- est precision score for passage type spoilers in Subtask 1. 1 Introduction The goal of SemEval-2023 Task 5 is to spoil click- bait. Clickbait posts are texts that arouse curiosity not by providing informative summaries of arti- cles, but by purposefully teasing and leaving out key information from an article to advertise a web page’s content. The aim of this shared task is to generate short texts (spoilers) that satisfy the cu- riosity induced by clickbait posts and provide more informative summaries of the linked articles. Al- though the first subtask may seem unnecessary, our results suggest that it could be highly helpful for the main purpose of spoiler generation. As click- bait posts are getting more common every day, this task holds more importance than ever. This task is only aimed towards English clickbait posts (Fröbe et al., 2023a). Our strategy for this shared task was to approach it as a question-answering problem. Our sys- tem uses the state-of-the-art sequence-to-sequence (Seq2Seq) model T5 (Raffel et al., 2019). We exper- iment with different variants of T5, some of which were previously fine-tuned for question-answering tasks. We implement a different model for each spoiler category type due to the contrast between the different types of output sequences. Our system achieves competitive results in both subtasks compared to the baseline from the shared task organizers and to the other teams. It has the highest precision for classifying the passage type spoilers in the spoiler classification subtask, and ranks as the top submission for phrase type spoilers in the generation subtask. However, our system struggles with generating longer spoilers such as phrase and multi type spoilers. We submitted our system as a docker image on the TIRA platform (Fröbe et al., 2023b) to increase reproducibility. We release our code 1 and the model checkpoints 2. 2 Background 2.1 Related Work A limited number of studies have directly focused on the task of clickbait spoiling. The most exten- sive work comes from Hagen et al. (2022), the organizers of this shared task. In their study, they approach clickbait spoiling as a question-answering and passage …

Content truncated to 500 words