An Integrated Graph of Words with Graph Neural Network for Cross-Lingual Classification to Detect Social Network Toxic Comments

With the popularity of social networks and the rapid growth of information on this platform, detecting and classifying toxic comments is necessary to protect social network users from fraudulent or potentially harmful activities. Because toxic comments on social networks are often in the form of very short lengths, applying traditional machine learning methods could not be highly effective, so a new rational approach is needed. This also motivates us to propose a new model called TextGCN4CLC using deep learning to classify cross-lingual short texts, based on the word-document graph approach and the graph neural network GCN, a new system in natural language processing for detecting social network toxic comments. By constructing word-document graphs, we model the semantic relationships between word-word, word-document, and document-document in a graph structure. Our proposed model TextGCN4CLC automatically detects toxic comments on social media. The results of extensive tests in the standard dataset show that TextGCNCLC achieves higher performance than state-of-the-art models.