Entity Role Classification through Contextual Embeddings and Aspect-Based Sentiment Features

Abstract

Propaganda is becoming increasingly influential on people’s everyday lives, partially due to recent developments such as the growth of social media. It has played a role in many recent global events, raising concerns. This phenomenon has called for ef- forts in the field of Natural Language Understanding (NLU) to study the language used in propaganda, and to develop methods that automatically detect propaganda to mitigate its spread. Propaganda is a complex phenomenon that changes with the time, and can sometimes be difficult to distinguish from similar phenomena such as persuasion. Furthermore, various techniques are used in propaganda, such as emotion manipulation and framing. Past research has focused on such techniques, but there has not been a lot of research on entity framing specifically, which is the focus of this study. Specifically, the effect of contextual embeddings and Aspect-Based Senti- ment Analysis was studied on entity role classification. The results show that ABSA improves the results of Support Vector Machine (SVM) models. This finding is an important first step towards entity classification in propaganda, which in turn is an important step towards propaganda detection and understanding in general. Future studies should focus on incorporating more propaganda techniques and their aspects into entity classification, and on classification of fine-grained roles.