TWEETSPIN: Fine-grained Propaganda Detection in Social Media Using Multi-View Representations

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3433 - 3448 July 10-15, 2022 ©2022 Association for Computational Linguistics TWEETSPIN: Fine-grained Propaganda Detection in Social Media Using Multi-View Representations Prashanth Vijayaraghavan IBM Research Almaden Lab San Jose, CA prashanthv@ibm.com Soroush Vosoughi Dartmouth College Department of Computer Science Hanover, NH soroush@dartmouth.edu Abstract Recently, several studies on propaganda de- tection have involved document and fragment- level analyses of news articles. However, there are significant data and modeling chal- lenges dealing with fine-grained detection of propaganda on social media. In this work, we present TWEETSPIN, a dataset containing tweets that are weakly annotated with different fine-grained propaganda techniques, and pro- pose a neural approach to detect and categorize propaganda tweets across those fine-grained categories. These categories include specific rhetorical and psychological techniques, rang- ing from leveraging emotions to using logical fallacies. Our model relies on multi-view repre- sentations of the input tweet data to (a) extract different aspects of the input text including the context, entities, their relationships, and exter- nal knowledge; (b) model their mutual inter- play; and (c) effectively speed up the learning process by requiring fewer training examples. Our method allows for representation enrich- ment leading to better detection and categoriza- tion of propaganda on social media. We verify the effectiveness of our proposed method on TWEETSPIN and further probe how the implicit relations between the views impact the perfor- mance. Our experiments show that our model is able to outperform several benchmark meth- ods and transfer the knowledge to relatively low-resource news domains. 1 Introduction Propaganda refers to any idea or information, that is often false or exaggerated, and is used to promote or publicize a particular cause or point of view. In recent years, there has been a surge in research and development of methods to detect propaganda from text. For example, some of the earlier works like (Rashkin et al., 2017a) and (Barrón-Cedeno et al., 2019) released a corpus of news articles containing coarse-grained document-level annotation of pro- paganda. Da San Martino et al. (2019b) described a corpus of news articles containing annotations of 18 fine-grained propaganda techniques. Following this work, two subtasks were presented as a part of NLP4IF workshop (Da San Martino et al., 2019a) that focused on the identification of propagandist text units at fragment and sentence level. We believe that the challenge for this task lies in the varied nature of propaganda techniques, includ- ing cognitive and information distortion and logical fallacies. The challenges are further exacerbated in the social media setting, which has become a key battleground in the spread of propaganda. Research on propaganda detection in social media platforms like Twitter has been limited by the: (a) lack of suf- ficiently annotated social media propaganda data, (b) idiosyncratic nature of the content on social media, (c) difficulty in modeling the social con- text in which propaganda is disseminated, and (d) varying propaganda techniques that require factual knowledge, structural relationships, …

Content truncated to 500 words