UMUTeam at SemEval-2023 Task 3: Multilingual transformer-based model for detecting the Genre, the Framing, and the Persuasion Techniques in Online News

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 609–615 July 13-14, 2023 ©2023 Association for Computational Linguistics UMUTeam at SemEval-2023 Task 3: Multilingual transformer-based model for detecting the Genre, the Framing, and the Persuasion Techniques in Online News Ronghao Pan1, José Antonio García-Díaz1, Miguel Ángel Rodríguez-García2, Rafael Valencia-García1 1 Facultad de Informática, Universidad de Murcia, Campus de Espinardo, 30100 Murcia, Spain {ronghao.pan@um.es, joseantonio.garcia8, valencia}@um.es 2Departamento de Ciencias de la Computación, Universidad Rey Juan Carlos, 28933 Madrid, Spain miguel.rodriguez@urjc.es Abstract In this manuscript, we describe the participa- tion of the UMUTeam in SemEval-2023 Task 3, a shared task on detecting different aspects of news articles and other web documents, such as document category, framing dimensions, and persuasion technique in a multilingual setup. The task has been organized into three related subtasks, and we have been involved in the first two. Our approach is based on a fine-tuned mul- tilingual transformer-based model that uses the dataset of all languages at once and a sentence transformer model to extract the most relevant chunk of a text for subtasks 1 and 2. The in- put data was truncated to 200 tokens with 50 overlaps using the sentence-transformer model to obtain the subset of text most related to the articles’ titles. Our system has performed good results in subtask 1 in most languages, and in some cases, such as French and German, we have archived first place in the official leader board. As for task 2, our system has also per- formed very well in all languages, ranking in all the top 10. 1 Introduction This working note describes the participation of the UMUTeam in the SemEval 2023 shared task 3 (Piskorski et al., 2023), focused on identifying different aspects of news articles, including docu- ment category, framing dimensions, and persuasion technique in a multilingual setup. Online media monitoring in multiple languages is an essential capability for organizations and com- panies that analyze and report on specific topical issues in different countries. However, due to the large amount of data to be processed and the in- terest in comparing in depth how topics of interest are treated and presented in different countries and media, the automation of media analysis processes is necessary. Automating media analysis processes can help reduce time and effort while providing a more complete and detailed overview of media coverage in different languages and geographies. It can also help identify trends, patterns, and devi- ations that would otherwise be difficult to detect. Therefore, this task focuses on recognizing differ- ent complementary aspects of a persuasive text, such as genre and persuasive techniques used to influence the reader. Advances in deep learning have enabled the de- velopment of very sophisticated and deep machine learning models. These include transformer-based models, a neural network architecture that has a weighty impact on NLP, making them suitable for tasks such as text classification, automatic trans- lation, and sentiment analysis (Bozinovski, 2020). In addition, they are trained on large datasets and are able to …

Content truncated to 500 words