YNUtaoxin at SemEval-2020 Task 11: Identification Fragments of Propaganda Technique by Neural Sequence Labeling Models with Different Tagging Schemes and Pre-trained Language Model

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1875–1880 Barcelona, Spain (Online), December 12, 2020. 1875 YNUtaoxin at SemEval-2020 Task 11: Identification Fragments of Propaganda Technique by Neural Sequence Labeling Models with Different Tagging Schemes and Pre-trained Language Model Xin Tao, Xiaobing Zhou∗ School of Information Science and Engineering Yunnan University, Yunnan, P.R. China ∗Corresponding author, zhouxb@ynu.edu.cn Abstract Information extraction is a hot topic in NLP, and detecting the use of propaganda techniques in news articles is part of this kind of task. This paper describes the solution of the Span Identification subtask in the Semeval 2020 Task 11: Detection of Propaganda Techniques in News Articles. The core idea of our method is equivalent to regard this task as a sequence tagging task and develop a neural sequence model to solve it. We use three different tagging schemes to tag sentences. Some pre-trained language models are used as the feature encoder like BERT, RoBERTa, and XLNet. In the final evaluation, we achieve the F1-score of 0.43208 and rank 11th among all the submitted teams. 1 Introduction Propaganda is a form of information promoting or publicizing a particular cause or point of view, especially one of a biased or misleading nature. Propaganda usually uses psychological and rhetorical techniques to meet its purpose. In today’s information age, propaganda is widely used in politics, business, science and technology, and other fields. And there are increasing propaganda techniques that are becoming more diverse and difficult to identify (Glowacki et al., 2018). We can search for the use of specific propaganda techniques via analyzing the language of an article, which is useful to find out the aims of propagandist news articles. In the mass of propagandist news articles, how to quickly and accurately identify the use of propaganda techniques is a challenging task. SemEval-2020 Task 11 - Detection of Propaganda Techniques in News Articles aims to develop automatic tools to detect propaganda techniques. And there are two subtasks: 1. Span Identification (SI): Given a plain-text document, identify those specific fragments which contain at least one propaganda technique. 2. Technique Classification (TC): Given a text fragment identified as propaganda and its document context, identify the applied propaganda technique in the fragment. The shared task provides a news article corpus, as well as 18 predefined propaganda techniques that have been annotated correspondingly. More detailed information can be seen in (Da San Martino et al., 2020). The rest of the paper is divided into the following sections: Section 2 describes the previous works and the relationship between coarse-grained and fine-grained tasks. Section 3 reports the final approach and those we tried . Section 4 details the results for the experimented models followed by error analysis of the best model. 2 Related Work Some previous works have proposed various methods for propaganda detection. LSTM is used for propaganda detection (Rashkin et al., 2017). Barr´on-Cede˜no et al. (2019) used the Maximum Entropy classifier with different features to deal with the same work of (Rashkin et al., 2017). However, …

Content truncated to 500 words