Enhanced Deep Learning Model for Violence, Racist, and Harassment Chat Detection in Child-Friendly Online Game

Violence in child-friendly rated online games didn't only happen visually but also in chat text. There are violent, racist, or harassing chat texts that often occur when children play online games, and it is usually beyond the parents' supervision. This paper proposes a model to detect violence, racism, and harassment from in-game chat to help parents supervise. The proposed model uses deep learning techniques such as BERT and BiLSTM to build the violence detection model. BERT is used to extract the feature without leaving the semantic context of a sentence, and then BiLSTM will classify the chat into neutral, violent, racist, or harassing. The dataset that will be used is the Indonesian Chat Dataset, which is collected from the most played child-friendly rated online games, titled Roblox and Minecraft. The collection consists of around $\mathbf{1 0, 7 0 2}$ curated chats. The result of the proposed model achieved 94.30 % accuracy and surpassed some other models that do the same on social media or games.