Kb at SemEval-2023 Task 3: On Multitask Hierarchical BERT Base Neural Network for Multi-label Persuasion Techniques Detection

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1395–1400 July 13-14, 2023 ©2023 Association for Computational Linguistics Kb at SemEval-2023 Task 3: On Multitask Hierarchical BERT Base Neural Network for Multi-label Persuasion Techniques Detection Katarzyna Baraniak and Marcin Sydow Polish-Japanese Academy of Information Technology kbaraniak@pjwstk.edu.pl msyd@pjwstk.edu.pl Abstract This paper presents a solution for Semeval 2023 subtask3 of task3: persuasion techniques in paragraphs detection. The aim of this task is to identify all persuasion techniques in each paragraph of given news articles. We use hier- archical multitask neural networks combined with transformers. Span detection is an aux- iliary task that helps in the main task: iden- tify propaganda techniques. Our experiments show that if we change the index of BERT em- bedding from the first token of the whole input to the first token of the identified span it can improve the performance. Span and label de- tection can be performed using one network so we save data and, when data is limited, we can use more of it for training. 1 Introduction News articles are expected to present reliable infor- mation, however quite often contain some kind of manipulation. Unconscious reader can be unable to spot all kind of persuasion that he is exposed to. Reading such news can influence a reader’s point of view especially if the reader has a low level of political knowledge (Eberl et al., 2017). Automatic tools for detection of persuasion can help in media analysis and creating more objective news. The task prepared by the organisers (Piskorski et al., 2023) covers three topics, as follows: cate- gory of news article identification, framing classifi- cation and paragraph-level persuasion techniques recognition. Our team focused on the third topic and created solution tested on all languages that have training data available: English, French, Ger- man, Italian, Polish and Russian and for additional languages with only test data: Spanish, Greek and Georgian. Our system combines multitask learning and hi- erarchical neural networks. The system identifies span where the persuasion techniques appear and then passes this information to the next module that classifies it to one or more categories of persuasion techniques. We used pre-trained and fine-tuned Bert embedding as the shared input layer. We discover that adding an auxiliary task such as span identification as the first step of training neural network may give better results than simply classifying the whole paragraph. Moreover, using one neural network for that approach makes solving the task possible even for limited data. The code of our solution is available at https://github.com/Katarzynaa/ persuasion_detection 2 Background The persuasion technique identification problem has been addressed before as a problem of span and technique classification. There exist several works on solving the persuasion techniques detec- tion problem, for example: as a problem of multi- modal image and text input (Dimitrov et al., 2021) or as problem of span and technique classification (Da San Martino et al., 2020). In our case it is the multi-label classification task, the authors of the task …

Content truncated to 500 words