A deep semantic-aware approach for Cantonese rumor detection in social networks with graph convolutional network

The rapid development of social networks provides people with opportunities for communication, which also makes it easier for the spread of rumors. In addition to Mandarin Chinese and English rumors, Cantonese rumors have been a major concern on social networks. However, there is no available Cantonese rumor dataset that includes information of propagation structures. Moreover, existing approaches focused on Mandarin Chinese cannot be applied directly to Cantonese rumor detection because of the differences of words in glyphs and pronunciations between them. In this paper, we construct the first Cantonese rumor dataset with abundant propagation structure information. Moreover, a novel deep semantic-aware graph convolutional network is proposed for Cantonese rumor detection, which integrates the global structural information and the local semantic features of Cantonese posts. To be specific, a CantoneseBERT model is designed to encode deep semantic and syntactic representations of Cantonese text contents, which introduces Cantonese glyph and Jyutping embeddings into the inputs of the model. In addition, a Bi-GCN model is used to extract the propagation clues and dispersion information from two social network graphs with opposite directions. Experimental results demonstrate that the proposed model outperforms the state-of-the-art models with an F-score of 0.8686. •Build the first Cantonese rumor dataset with propagation structure information.•Propose a novel method to deeply mine the semantic information of Cantonese.•Design a novel graph-structured model SA-GCN for Cantonese rumor detection.•Conduct the first study to detect Cantonese rumors based on propagation structures.