Detecting Propaganda in News with Hierarchical Graphs

Social media has become a hotbed for spreading facts, false claims, and propaganda, especially since the Covid-19 pandemic. Graph Neural Networks (GNNs) are great at processing language, but detecting propaganda is tricky because of the complexity of text interactions and the need to understand context that's not right next to each other. Scientists have come up with a new method called Hierarchical Graph-based Integration Network (H-GIN) to tackle this problem.

H-GIN works by creating a two-layer graph. The first layer uses something called Residual-driven Enhancement and Processing (RDEP) to connect distant nodes of information. The second layer, called Attention-driven Multichannel feature Fusing (ADMF), merges different types of information—like sequences, meanings, and grammar—to detect propaganda. The model trains on existing datasets like ProText, Qprop, and PTC. When tested, H-GIN performed really well, achieving 82% accuracy. It could also spot propaganda in new, unseen cases with the same high accuracy, using the ProText dataset. This shows that H-GIN can be a reliable tool to identify propaganda on social media.

Actions