CoT-based Data Augmentation Strategy for Persuasion Techniques Detection
Dailin Li, Chuhan Wang, Xin Zou, Junlong Wang, Peng Chen, Jian Wang, Liang Yang, Hongfei Lin · 2024
Memes are commonly used in online disinformation campaigns, particularly on social media platforms.They are primarily effective on social media platforms since they can easily reach many users.Semeval2024-Task4 (Dimitrov et al., 2024), "Multilingual detection of persuasion techniques in memes", focuses on detecting persuasive methods across four languages: English, Bulgarian, North Macedonian and Arabic.Subtask 1 aims to identify the given text fragments of memes and which of the 20 persuasion techniques it uses, organized in a hierarchy.For the difficulty of this task and the fundamental role of text in the artificial intelligence area, we concentrate solely on this task.We develop a system using CoT-based data augmentation methods,in-domain pretraining and ensemble strategy that combines the strengths of both RoBERTa and DeBERTa models.Our solution achieved the top ranking among 33 teams in the English track during the official assessments.We also analyze the impact of architectural decisions, data construction and training strategies.We release our code at https://github.com/ldlbest/semeval2024-task4