ReDASPersuasion at SemEval-2023 Task 3: Persuasion Detection using Multilingual Transformers and Language Agnostic Features

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2124–2132 July 13-14, 2023 ©2023 Association for Computational Linguistics ReDASPersuasion at SemEval-2023 Task 3: Persuasion Detection using Multilingual Transformers and Language Agnostic Features Fatima Zahra Qachfar University of Houston Houston, TX, USA fqachfar@uh.edu Rakesh M. Verma University of Houston Houston, TX, USA rmverma2@central.uh.edu Abstract This paper describes a multilingual persuasion detection system that incorporates persuasion technique attributes for a multi-label classifi- cation task. The proposed method has two ad- vantages. First, it combines persuasion features with a sequence classification transformer to classify persuasion techniques. Second, it is a language agnostic approach that supports a total of 100 languages, guaranteed by the mul- tilingual transformer module and the Google translator interface. Our persuasion system out- performs the SemEval baseline in all languages, except zero shot prediction languages, which was not the main focus of our research. With the highest F1-Micro score of 0.45 for Italian, it achieved the eighth position on the leaderboard. 1 Introduction The goal of the SemEval-2023 Task 3 (Piskorski et al., 2023) Subtask 3 is to detect the presence of one or many persuasion techniques in text data. It is a multilingual task and consisted initially of six languages: English, French, Polish, Italian, Rus- sian and German. In the testing phase, organizers added three surprise languages (Spanish, Greek, and Georgian) for zero-shot classification tasks. In this work, we aim to build a cross-lingual pre- dictive model that detects persuasion techniques in a given text with the help of language transformers and linguistic features. Our system uses XLM- RoBERTa (Conneau et al., 2019) as a backbone model, incorporating language agnostic features provided using Google translation over 100 lan- guages for persuasion detection. XLM-RoBERTa has been shown to be a powerful multilingual pre- trained language model compared against other models like Multilingual BERT (M-BERT) (Devlin et al., 2018) and DistilBERT (Sanh et al., 2019). Also, this language transformer can process all the languages existing in the SemEval-2023 Task3. Through this task, we realized the importance of data quality in Multi-label classification tasks with cross-lingual settings. We were also able to build a new system that incorporates language agnostic features to capture persuasion in text semantics, accompanied by the powerful capacity of language transformers based on attention mechanism. 2 Background 2.1 Task Description The SemEval-2023 Task 3 (Piskorski et al., 2023) had three subtasks: (i) Subtask 1 is news genre cat- egorization, (ii) Subtask 2 is framing detection, and (iii) Subtask 3 is persuasion techniques detection. We focus solely on the multi-label classification problem Subtask 3 that aims to detect 23 persua- sion techniques. All subtasks are multilingual and contain zero-shot learning. In Subtask 3, each article in the dataset is labeled with multiple classes. One article can contain dif- ferent persuasion techniques (persuasion data statis- tics are listed in Table 6, Appendix B). 2.2 Task Evaluation Subtask 3 is evaluated using the Micro F1 score between the gold labels and model predictions de- fined in Equation (1). …

Content truncated to 500 words