Sabrina Spellman at SemEval-2023 Task 5: Discover the Shocking Truth Behind this Composite Approach to Clickbait Spoiling!

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 969–977 July 13-14, 2023 ©2023 Association for Computational Linguistics Sabrina Spellman at SemEval-2023 Task 5: Discover the Shocking Truth Behind this Composite Approach to Clickbait Spoiling! Simon Birkenheuer and Jonathan Drechsel and Paul Justen and Jimmy Pöhlmann and Julius Gonsior and Anja Reusch Technische Universität Dresden firstname.lastname@mailbox.tu-dresden.de Abstract This paper describes an approach to automat- ically close the knowledge gap of Clickbait- Posts via a transformer model trained for Question-Answering, augmented by a task- specific post-processing step. This was part of the SemEval 2023 Clickbait shared task (Fröbe et al., 2023a) - specifically task 5. We devised strategies to improve the existing model to fit the task better, e.g. with different special mod- els and a post-processor tailored to different inherent challenges of the task. Furthermore, we explored the possibility of expanding the original training data by using strategies from Heuristic Labeling and Semi-Supervised Learn- ing. With those adjustments, we were able to improve the baseline by 9.8 percentage points to a BLEU-4 score of 48.0%. 1 Introduction This project was part of the SemEval 2023 Click- bait shared task (Fröbe et al., 2023a). Specifically, we focused on the task of generating spoilers for clickbait posts. We viewed this task as an extrac- tive question-answering problem, since a knowl- edge gap (or question) posed by a headline shall be closed or answered by a passage from an arti- cle. Based on this assessment, a transformer-based approach has been chosen since these approaches tend to perform well for question-answering prob- lems (Shao et al., 2019). To improve the perfor- mance of our models, additional data for use in the training process has been generated. In ad- dition, multiple models have been trained using different approaches. Another challenge is that some of the examples require multiple spoilers to be generated as an answer. To address this issue, a post-processing approach to generate multiple relevant answers that are not too similar to each other has been developed. Furthermore, a specific pattern could be identified and used to develop a heuristic in the pre- and post-processing step to find solutions to these examples independent of the model. 2 Background The task we attempted to solve was as follows: given a post text that introduces a so-called knowl- edge gap that piques the reader’s interest in read- ing the accompanying article, our objective was to identify the passages within this article that best addresses this knowledge gap. 2.1 Clickbait Dataset A dataset was provided for solving the task (Ha- gen et al., 2022) to which we will refer to as the "Clickbait Dataset". This dataset consists of 3200 training samples and 800 validation samples. An- other 1000 test samples were withheld and were only accessible through the evaluation portal TIRA. Each of these samples contains a postText which is the text of a clickbait post. A clickbait post refers to an article that shall be read. Therefore, the title and also the text …

Content truncated to 500 words