language:
- cs
license: cc-by-nc-sa-4.0
size_categories:
- 1K<n<10K
dataset_info:
features:
- name: id
dtype: string
- name: text
dtype: string
- name: genre
dtype: string
- name: topic
dtype: string
- name: scope
dtype: string
- name: location
dtype: string
- name: argumentation
dtype: string
- name: emotions
dtype: string
- name: overall_sentiment
dtype: string
- name: russia
dtype: string
- name: opinion
dtype: string
- name: expert
dtype: string
- name: source
dtype: string
- name: fear-mongering
dtype: string
- name: blaming
dtype: string
- name: labeling
dtype: string
- name: demonization
dtype: string
- name: relativization
dtype: string
- name: fabulation
dtype: string
- name: ranges
list:
- name: attribute
dtype: string
- name: end
dtype: int64
- name: start
dtype: int64
- name: text
dtype: string
splits:
- name: train
num_bytes: 27173943
num_examples: 7642
- name: test
num_bytes: 3727325
num_examples: 1000
download_size: 19285049
dataset_size: 30901268
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Dataset Card for the benchmark Propaganda Dataset
Propaganda corpus is a joint work between multiple faculties of Masaryk University (Faculty of Social Sciences, Faculty of Informatics, and Faculty of Law) under the project Manipulative techniques of propaganda in the age of Internet. In its current state, the dataset contains 8,646 documents that were extracted from four Czech news websites. These websites were previously investigated for distributing Russian propaganda.
Dataset Details
Each document is annotated with three types of attributes:
- Manipulative techniques:
- relate to specific sections of the document
Attribute Classes Description Argumentation yes, no Does the text present facts or arguments (logical, emotional, etc.) to support the main claim? Blaming yes, no Does the text accuse someone of something? Demonization yes, no Is the “enemy” and/or his/her goals or interests presented in the text as being evil Emotions grieviance, hatred, compassion, fear, missing What is the main emotion the text is trying to evoke in the reader? Fabulation yes, no Does the text contain unsubstantiated, overstated or otherwise incorrect claims? Fear-mongering yes, no Is the text trying to appeal to fear, uncertainty or other threat? Labeling yes, no The text uses specific labels – short and impactful phrases or words – to describe a person, group or object. Relativization yes, no Are the presented actions of a person, group or party being relativized?
Global attributes:
Attribute Classes Description Genre news, comment, interview The publication form of the news text. Location EU, Czech Republic, USA, Russia, NATO, Russia + USA, other locations, other/cannot be determined What is the main location the text discusses about? Overall Sentiment positive, negative, neutral The core sentiment of the newspaper text. Topic migration crisis, domestic politics, foreign policy / diplomacy, society / social situation, energy, economy / finance, conflict in Ukraine, conflict in Syria, conspiracy, other, culture, social policy, arms policy various topics Scope foreign, domestic, both, cannot be determined Distinguishes domestic and foreign topics Other attributes:
- do no fit into any other categories (they relate to a specific section of a document but are not manipulative techniques by themselves)
Attribute Classes Description Expert yes, no Is the text or opinion in the text presented as being supported by an expert? Opinion yes, no Does the author of the text present his or her personal opinion? Russia positive example, neutral, victim, negative example, hero, missing How Russia is depicted in the article? Source yes, no Is the text presented as being based on a specific source?
Dataset Description
The benchmark Propaganda dataset contains 8,646 newspaper articles from 2016 (5,500 documents, 2,7 million tokens), 2017 (1,994 documents, 930 thousand tokens), and 2018 (1,152 documents, 500 thousand tokens). Compared with other resources, the Propaganda dataset contains fine-grained annotations of both document-level attributes and specific text devices exemplified by marked phrases from the article texts.
The Czech Republic was selected here as a representative of a country within the former Soviet Union influence and, as such, with significantly active propaganda sources. The analyzed news texts were downloaded from four newspaper media outlets publishing in the Czech language:
- Sputnik News
- Parlamentní listy (Parliamentary Letters)
- AC24
- Svět kolem nás (The World around Us).
Dataset Sources
- Repository: https://nlp.fi.muni.cz/projects/propaganda/dataset/
- Paper: https://dx.doi.org/10.1016/j.eswa.2024.124085
- Demo: https://nlp.fi.muni.cz/projects/propaganda
Citation
BibTeX:
@article{horak_etal2024_recognition,
title = {Recognition of propaganda techniques in newspaper texts: Fusion of content and style analysis},
author = {Aleš Horák and Radoslav Sabol and Ondřej Herman and Vít Baisa},
journal = {Expert Systems with Applications},
pages = {124085},
year = {2024},
issn = {0957-4174},
publisher = {Elsevier},
doi = {https://doi.org/10.1016/j.eswa.2024.124085},
}
APA:
Aleš HORÁK, Radoslav SABOL, Ondřej HERMAN and Vít BAISA. Recognition of Propaganda Techniques in Newspaper Texts: Fusion of Content and Style Analysis. Expert Systems with Applications. Elsevier, 2024. ISSN 0957-4174. https://dx.doi.org/10.1016/j.eswa.2024.124085.