CNLP-NITS at SemEval-2023 Task 10: Online sexism prediction, PREDHATE!

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 88–93 July 13-14, 2023 ©2023 Association for Computational Linguistics Mr-Fosdick at SemEval-2023 Task 5: Comparing Dataset Expansion Techniques for Non-Transformer and Transformer Models: Improving Model Performance through Data Augmentation Christian Falkenberg and Erik Schönwälder and Tom Rietzke Chris-Andris Görner and Robert Walther and Julius Gonsior and Anja Reusch Technische Universität Dresden first.last@mailbox.tu-dresden.de Abstract In supervised learning, a significant amount of data is essential. To achieve this, we gener- ated and evaluated datasets based on a provided dataset using transformer and non-transformer models. By utilizing these generated datasets during the training of new models, we attain a higher balanced accuracy during validation compared to using only the original dataset. 1 Introduction Titles of posts are increasingly utilizing exagger- ations and other persuasive techniques to entice clicks. Such titles like the following are known as clickbait: ’Stranger Knocked Angrily on Her Door – What He Left Made Her Run to the Phone’ The Clickbait Challenge, hosted on Tira (Fröbe et al., 2023b), is a shared task that aims to address and satisfy the curiosity generated by this phenomenon. The challenge comprises two subtasks that utilize the post titles featuring clickbait and their corresponding articles as input. In the first task, three types of spoilers are to be distinguished, while the second task involves retrieving the spoiler from the article. For a detailed description of the tasks, please refer to the Overview Paper (Fröbe et al., 2023a) or the official Clickbait Challenge website (Clickbait). We concentrated on optimising the given dataset to provide better training data for the models we created. In the following, we will describe our ap- proaches to obtaining new datasets. We then evalu- ate our generated datasets and compare them to the original datasets. Our approaches for task 1, click- bait classification, using a non-transformer and a transformer-based approach are then described. Af- ter that we present our model for task 2 and show how we retrieve the spoilers for the clickbait titles. 2 Expanding the Dataset 2.1 Dataset Generation The original dataset provided to us (Hagen et al., 2022) contained 3200 examples of training data. All examples consisted of a clickbait post, the ar- ticle text, the spoiler type and the corresponding spoiler. Table 1 presents an overview of all utilized and created datasets, including their corresponding source file sizes. 2.1.1 QA Dataset Our initial approach was to use existing data and format it for our purposes. For example, question answering datasets (QA) are appropriate for the task at hand. In QA datasets, the question can be understood as PostText and the answer as a spoiler. The length of the spoiler then in turn determines the spoiler type. To create our own suitable MixQA (MixQA Dataset) dataset, several QA datasets 1 were combined. The largest and best known dataset in this series is the SQuAD2.0 (Rajpurkar et al., 2018) dataset. 2.1.2 Pegasus Dataset Data generation through paraphrasing is a process where new, diverse and meaningful samples …

Content truncated to 500 words