newsSweeper at SemEval-2020 Task 11: Context-Aware Rich Feature Representations for Propaganda Classification

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1764–1770 Barcelona, Spain (Online), December 12, 2020. 1764 newsSweeper at SemEval-2020 Task 11: Context-Aware Rich Feature Representations For Propaganda Classification Paramansh Singh∗ Siraj Sandhu∗ Subham Kumar∗ Ashutosh Modi Indian Institute of Technology Kanpur (IITK) {params,sssandhu,subhamkr}@iitk.ac.in ashutoshm@cse.iitk.ac.in Abstract This paper describes our submissions to SemEval 2020 Task 11: Detection of Propaganda Techniques in News Articles for each of the two subtasks of Span Identification and Technique Classification. We make use of pre-trained BERT language model enhanced with tagging tech- niques developed for the task of Named Entity Recognition (NER), to develop a system for identifying propaganda spans in the text. For the second subtask, we incorporate contextual features in a pre-trained RoBERTa model for the classification of propaganda techniques. We were ranked 5th in the propaganda technique classification subtask. 1 Introduction Propaganda is commonly defined as information of a biased or misleading nature, possibly purposefully shaped, to promote an agenda or a cause. Commonly used propaganda techniques are psychological and rhetorical - ranging from the selective presentation of facts and logical fallacies to the use of loaded language to produce an emotional response. Historically, propaganda has been widely employed and is often associated with governments, activists, big business, religious organizations, and partisan media. Having been first introduced in a previous iteration of the SemEval shared task (Martino et al., 2019), the current shared task (Da San Martino et al., 2020) consists of two subtasks, (i) Span Identification(SI): Given a plain-text document, identify those specific fragments which are propagandistic, and (ii) Tech- nique Classification(TC): Given a text fragment identified as propaganda and its document context, identify the applied propaganda technique in the fragment. Both the subtasks focus on English texts only. An illustration is shown in Figure 1. Figure 1: Example of subtasks span identification(SI) and technique classification(TC) The previous iteration of this task required identifying the propaganda spans and also identifying the propaganda technique for that fragment. The highest scoring system by Yoosuf and Yang (2019) employed an out-of-the-box BERT (Devlin et al., 2018) language model fine-tuned on the token classification task - whether the token is part of a propaganda span or not. For the classification task, the same system employed an N+2 way classification, where N is the number of propaganda classes. For the technique ∗ Authors equally contributed to this work. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/. 1765 classification, we make use of document context. For context inclusion, we build upon an approach by Hou and Chen (2019), where they use title and previous sentence as context. They pass the context concatenated with the original text for fine-tuning the BERT model. Our approach for span identification makes use of a state-of-the-art language model enhanced by tagging schemes inspired from Named Entity Recognition which aim to model spans better - logically and intuitively - by involving Begin and End tags (Ramshaw and Marcus, 1995) to better formulate spans. In …

Content truncated to 500 words