Purpose A large part of the misinformation, fake news, and propaganda spread on social media originates from content disseminated via online social network platforms, such as X (formerly Twitter) and Facebook. The control and filtering of digital media pose significant challenges and threats to online social networking. This paper aims to understand how propaganda infiltrates news articles, which is critical for fully grasping its impact on daily life.
Design/methodology/approach This study introduces a pre-trained language model framework, called ProST, to detect propaganda in text-based news articles. ProST addresses two tasks: identifying propaganda spans and classifying propaganda techniques. For span identification, we built a model combining a pre-trained RoBERTa model with long-short-term memory and begin, inside, outside and end tagging to detect propaganda spans. The technique classification model uses contextual features and a RoBERTa-based approach. This study, conducted on the SemEval-2020 dataset (comprising 536 news articles), demonstrates a performance comparable to state-of-the-art methods.
Findings The results indicate that the ProST model is highly effective in detecting propaganda in text news articles, accurately identifies propaganda spans and classifies techniques with high precision, benefitting from sentence- and span-level feature pruning.
Originality/value The ProST model offers a novel approach to identifying propaganda in online news articles with diverse webs of information. To the best of our knowledge, this is the first framework capable of classifying both propaganda spans and techniques in textual news. Accordingly, ProST represents a significant advancement in the field of propaganda.