Perceiving the Narrative Style for Fake News Detection Using Deep Learning

Existing deep-learning-based features have shown strong enough results (more than 90% accuracy) if a large amount of annotated data is available. However, in reality, data annotation is labor-intensive and expensive. This work proposes a novel approach for fake news detection by perceiving narrative style with deep learning to alleviate the problem. Deep-learning-based features are represented as embedding vectors. Traditional embedding vectors ingest word context information, but the training requires considerable annotated datasets. Many linguistics studies have shown that written styles, such as the usage of punctuation, repetition of words, and grammatical order, are significant for distinguishing fake news. Our model takes advantage of the word-to-word dependency relationship, describing the styles of the news utterances. We denote the proposed model as Syntax Graphical Thread (SGT) network. We utilize a trainable randomly initialized embedding and Gated Recurrent Unit (GRU) layer to capture the context vector, while a Graph Attention (GAT) layer is used to capture the narrative features. The experimental result manifests that our method can significantly mitigate the reliance on data scale and present better classification results when the dataset is limited