Text Moderation in Online Communities: Integrating User Attributes and Interaction Graph Embedding

Text moderation in online communities has always been an important part of preventing cybercrime and maintaining a clear cyberspace. Most existing text moderation research approaches it from a text classification perspective, focusing mainly on the text content itself with scant attention to the factors of the text publishers. Understanding user interaction in text moderation is an important part of engaging users to grow online communities. The paper presents a text moderation model for online communities, which is based on a text pre-trained model and integrates user interaction features and attribute features. Since it is difficult to label users in large online communities, we use the unsupervised Node2Vec algorithm to get user embedding from the user's interaction graph. Then the attention mechanism is used to further train the user embedding to promote the enhancement effect of user characteristics on text moderation. The experimental results on real community datasets show that the proposed method outperforms the text moderation models that only use text features. Simultaneously, from the perspective of community owner, we conduct numerous experiments to demonstrate the impact of community user attributes and interaction characteristics on text moderation, providing valuable insights for community managers.