Leveraging Declarative Knowledge in Text and First-Order Logic for Fine-Grained Propaganda Detection

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3895–3903, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3895 Leveraging Declarative Knowledge in Text and First-Order Logic for Fine-Grained Propaganda Detection Ruize Wang1∗, Duyu Tang2, Nan Duan2,Wanjun Zhong3, Zhongyu Wei1†, Xuanjing Huang1, Daxin Jiang2, Ming Zhou2 1Fudan University, Shanghai, China 2Microsoft Corporation, Beijing, China 3Sun Yat-Sen University, Guangzhou, China {rzwang18,zywei,xjhuang}@fudan.edu.cn {zhongwj25}@mail2.sysu.edu.cn {dutang,nanduan,djiang,mingzhou}@microsoft.com Abstract We study the detection of propagandistic text fragments in news articles. Instead of merely learning from input-output datapoints in train- ing data, we introduce an approach to inject declarative knowledge of fine-grained propa- ganda techniques. Specifically, we leverage the declarative knowledge expressed in both first-order logic and natural language. The for- mer refers to the logical consistency between coarse- and fine-grained predictions, which is used to regularize the training process with propositional Boolean expressions. The latter refers to the literal definition of each propa- ganda technique, which is utilized to get class representations for regularizing the model pa- rameters. We conduct experiments on Propa- ganda Techniques Corpus, a large manually annotated dataset for fine-grained propaganda detection. Experiments show that our method achieves superior performance, demonstrating that leveraging declarative knowledge can help the model to make more accurate predictions. 1 Introduction Propaganda is the approach deliberately designed with specific purposes to influence the opinions of readers. Different from the fake news which is entirely made-up and has no verifiable facts, pro- paganda is possibly built upon an element of truth, and conveys information with strong emotion or somewhat biased. This characteristic makes pro- paganda more effective and unnoticed through the rise of social media platforms. Some examples of propagandistic texts and definitions of correspond- ing techniques are shown in Figure 1. We study the problem of fine-grained propa- ganda detection in this work, which is possible Asia.∗Work is done during internship at Microsoft Research †Corresponding author. 1. Appeal to fear: Seeking to build support for an idea by instilling anxiety. �Our convention occurs at a moment of crisis for our nation1,� said Trump. “The attacks on our police, and the terrorism in our cities, threaten our very way of life1.� 3. Loaded language: Using specific words and phrases with strong emotional implications to influence an audience. “To all Americans tonight, in all our cities and towns, I make this promise: We Will Make America Strong Again. We Will Make America Proud Again. We Will Make America Safe Again. And We Will Make America Great Again 3, 4.” 4. Slogan: A brief and striking phrase that may include labeling and stereotyping. 2. Repetition: Repeating the same message over and over again. Trump tweeted: “I’m building a wall, OK? I’m building a wall2. I am going to do very well with the Hispanics, the Mexicans.” Figure 1: Examples of propagandistic texts, and defi- nitions of corresponding propaganda techniques (Bold denotes propagandistic texts). thanks to the recent release of Propaganda Tech- niques Corpus (Da San Martino et al., 2019). Dif- ferent from earlier works (Rashkin et al., …

Content truncated to 500 words