Appeal for Attention at SemEval-2023 Task 3: Data augmentation extension strategies for detection of online news persuasion techniques

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 616–623 July 13-14, 2023 ©2023 Association for Computational Linguistics Appeal for attention at SemEval-2023 Task 3: Data augmentation and extension strategies for detection of online news persuasion techniques Amihaesei Sergiu Cornei Laura Faculty of Computer Science "Alexandru Ioan Cuza" University of Iasi, Romania {amihaeseisergiu,cornei.laura10,sgeorge.sstoica99}@gmail.com Stoica George Abstract In this paper, we proposed and explored the im- pact of four different dataset augmentation and extension strategies that we used for solving the subtask 3 of SemEval-2023 Task 3: multi- label persuasion techniques classification in a multi-lingual context. We consider two types of augmentation methods (one based on a mod- ified version of synonym replacement and one based on translations) and two ways of extend- ing the training dataset (using filtered data gen- erated by GPT-3 and using a dataset from a previous competition). We studied the effects of the aforementioned techniques by using the augmented and/or extended training dataset to fine-tune a pretrained XLM-RoBERTa-Large model. Using the augmentation methods alone, we managed to obtain 3rd place for English, 13th place for Italian and between the 5th to 9th places for the other 7 languages during the competition. 1 Introduction The subtask 3 of SemEval-2023 Task 3 (Piskorski et al., 2023) focuses on detecting the persuasion techniques used in online news, in a multilingual context. This task is significantly important as it may enable one to identify attempts of manipu- lation and misinformation, attacks on reputation, persuasive language and logical fallacies in textual data. The data offered by the organizers of the com- petition covered nine languages: English, French, Italian, Spanish, German, Georgian, Greek, Polish and Russian. For three of the languages (Span- ish, Georgian and Greek) only the test data was available, to promote the development of language- agnostic solutions in a zero-shot learning scenario. To solve this subtask, we used a pre-trained multi-lingual transformer-based model: XLM- RoBERTa-Large (Conneau et al., 2020). We further fine-tuned this model on an augmented version of the dataset which encompassed the provided train- ing data for the six languages. We used two augmentation techniques that had the purpose to balance the amount of data associ- ated either with each persuasion technique or with each language. The first augmentation method con- sisted in creating new articles by translating the data from one language to all of the other languages using a translation API. The second technique in- volved employing a special procedure to balance the data corresponding to each persuasion tech- nique; in this case, the augmentation consisted in creating similar articles by replacing words from the paragraphs with synonyms from the NLTK WordNet corpus 1. By using data augmentation, we managed to score 3rd on the English language, and between 5 − 13th places on all the other languages on the official SemEval-2023 test set leaderboard. Post-competition, we considered extending the training dataset in two ways: by using a dataset from a similar previous competition (Dimitrov et al., 2021) and by using data …

Content truncated to 500 words