In the winter semester of 2022/2023, PhD students of the ICT&Design Doctoral School, Arkadiusz Modzelewski and Witold Sosnowski, qualified for the 17th edition of the International Workshop on Semantic Evaluation SemEval 2023. They participated in the task on SemEval-2023 Task 3: Detecting Category, Framing, and Persuasion Techniques in online news in a multilingual setting.
As part of the task, they investigated the ability of natural language processing (NLP) systems to detect genres and persuasion techniques across multiple languages and conducted experiments with various data augmentation techniques, including machine translation and text generation.
To this end, they used the OpenAI GPT-3 Davinci language model to create synthetic texts for genre detection, and for persuasion techniques detection, they relied on dataset expansion through text translation using the DeepL translator.
The fine-tuned models that used the augmented data ranked in the top ten across all languages, indicating the effectiveness of the proposed approach.
The genre detection models demonstrated excellent results, securing first, second, and third place in Spanish, German, and Italian, respectively.
Meanwhile, one of the persuasion technique detection models secured third position in Polish. Their contribution to science includes a system architecture that utilizes DeepL and GPT-3 for data augmentation to detect both genre and persuasion techniques.
The proposed methods have been described in the article: “DSHacker at SemEval-2023 Task 3: Genres and Persuasion Techniques Detection with Multilingual Data Augmentation through Machine Translation and Text Generation”, which has been accepted for publication. The SemEval workshop, where the competition results will be presented, will take place as part of the ACL-2023 conference in Toronto, Canada.
