Mr-wallace at SemEval-2023 Task 5: Novel Clickbait Spoiling Algorithm Using Natural Language Processing

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1625–1629 July 13-14, 2023 ©2023 Association for Computational Linguistics Mr-wallace at SemEval-2023 Task 5: Novel Clickbait Spoiling Algorithm Using Natural Language Processing Vineet Saravanan Cranbrook Schools Bloomfield Hills, MI vineetsaravanan@gmail.com Steven Wilson Oakland University Rochester, MI stevenwilson@oakland.edu Abstract Clickbait creates a nuisance in the online expe- rience by creating a lure towards poor content in order to generate ad revenue. With the use of natural language processing models, we can save users time and reduce the need to follow clickbait links. Task 5 at SemEval-2023 fo- cused on precisely this problem and was broken into two steps: identifying the clickbait spoiler type and then identifying the clickbait itself. Our approach involves the use of fine-tuned text classification and question-answering mod- els. Our classification model is able to deter- mine the type of clickbait with 65.3% accu- racy. The question-answering model exactly spoiled clickbait generated around 42.5% of the time. Efforts toward solving this task may have an impact by helping to save users’ time and quickly give an insight into the answer of what the clickbait/article is about. 1 Introduction Clickbait, in social media, are exaggerated head- lines whose main motive is to mislead the reader to “click” on them. They create a nuisance in the online experience by creating a lure towards poor content. Online content creators are utilizing more of them to get increased page views and, thereby, more ad revenue without providing the backing content (Agrawal, 2016). Clickbait articles usually appeal to a user’s emotions and curiosity. These articles are often lacking in quality and information accuracy. Clickbait articles have a considerably higher click rate than non-clickbait articles. This has been a rising problem in the 21st century with the growth in internet usage and social media. In SemEval-2023, the tasks of clickbait spoiling task is split into two subtasks, as shown in Figure 1, the first task is to identify the clickbait type, and the second task is to actually spoil the clickbait by identifying the sequence(s) of text in the source article that provide the information that the reader is being lured to read about given a particular click- bait post (Fröbe et al., 2023a). For Task 1, which is the Spoiler Type Classification, the input is the clickbait post and the linked document, and the task is to classify the spoiler type that the clickbait post warrants (either "phrase," "passage," or "multi"), which can be approached using a supervised text classification approach. For Task 2, which is on Spoiler Generation, the input is the clickbait post and the linked document. The task is to gener- ate the spoiler for the clickbait post, which can be approached using an information extraction or question-answering approach where the question is in the title of the post. More information about the source datasets can be found in Figure 2. We fine-tuned a DistilBERT (Sanh et al., 2019) model for task 1, which was one of the …

Content truncated to 500 words