TohokuNLP at SemEval-2023 Task 5: Clickbait Spoiling via Simple Seq2Seq Generation and Ensembling

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1756–1762 July 13-14, 2023 ©2023 Association for Computational Linguistics TohokuNLP at SemEval-2023 Task 5: Clickbait Spoiling via Simple Seq2seq Generation and Ensembling Hiroto Kurita1†, Ikumi Ito1†, Hiroaki Funayama1,2, Shota Sasaki3,2‡, Shoji Moriya1, Ye Mengyu1, Kazuma Kokuta1, Ryujin Hatakeyama1, Shusaku Sone1, Kentaro Inui1,2 1Tohoku University 2RIKEN AIP 3CyberAgent, Inc. {hiroto.kurita,ikumi.ito.p8,h.funa,shoji.moriya.q7,ye.mengyu.s1, kokuta.kazuma.r3,hatakeyama.ryujin.q7,sone.shusaku.r8}@dc.tohoku.ac.jp sasaki_shota@cyberagent.co.jp kentaro.inui@tohoku.ac.jp Abstract This paper describes our system submitted to SemEval-2023 Task 5: Clickbait Spoiling. We work on spoiler generation of the subtask 2 and develop a system which comprises two parts: 1) simple seq2seq spoiler generation and 2) post-hoc model ensembling. Using this simple method, we address the challenge of generating multipart spoiler. In the test set, our submitted system1 outperformed the baseline by a large margin (approximately 10 points above on the BLEU score) for mixed types of spoilers. We also found that our system successfully handled the challenge of the multipart spoiler, confirm- ing the effectiveness of our approach. 1 Introduction Clickbait is a type of text on social media that is specifically designed to exploit users’ curiosity and lure them into clicking on a linked webpage. Due to the deceptive and misleading nature of the content often presented, clickbait is generally considered harmful (Hagen et al., 2022). Thus addressing the clickbait issue is a crucial concern for social media users. Subtask 2 of SemEval-2023 Task 5 aims to gener- ate spoilers of clickbait based on posted text and the linked webpage (Fröbe et al., 2023a). Hagen et al. (2022) proposed a method for generating spoilers via encoder-based question answering approach. However, their method has limitations in handling a specific type of spoiler where the answer is spread across several sections of the target article. To tackle this problem, we develop a simple yet effective sequence-to-sequence (seq2seq) ap- proach for the subtask 2: spoiler generation. In our †equal contributions ‡Work done while at RIKEN AIP and Tohoku University. 1The docker image of our final submission on TIRA(Fröbe et al., 2023b) is registry.webis.de/code-research/ tira/tira-user-john-king/finalsub:0.0.4 and the implementation is available at https://github.com/ cl-tohoku/semeval2023-task5-seq2seq-ensemble. Figure 1: Overview of SemEval-2023 Task 5: Click- bait Spoiling. The task is to extract one or more spoilers for postText from content (targetTitle and targetParagraphs). The text highlighted with yellow in the content is the spoiler to be extracted. approach, the target spoiler is directly generated from the seq2seq model, rather than predicting its spans in the context. This enables us to handle the challenge of multipart spoiler with ease without implementing more complex approaches such as multiple-span extraction (Segal et al., 2020). Exper- imental results demonstrate that our approach suc- cessfully addresses the multipart spoilers, leading to an overall score improvement without degrading the scores of single-part spoilers. Furthermore, we incorporate a post-hoc ensemble of models with multiple seeds in our system, leading to enhanced performance over a single model. In analysis, we show that our system generates spoilers extractively rather than abstractively, even though we adapt the seq2seq generation approach. We …

Content truncated to 500 words