Abstract : This study addresses critical limitations in an existing profanity-filtering algorithm: insufficient context interpretation and absence of leetspeak detection. First, the researchers integrated the BERT transformer model to improve context-sensitive filtering, achieving a 99.1% accuracy rate and a 10.4% increase in correctly censoring 1,000 chat results from the Minecraft-Server-Chat dataset. Toxicity scoring with Toxic-BERT allowed precise filtering, distinguishing between friendly and harmful content words. Second, the researchers incorporated a reverse mapping function to identify leetspeak, significantly improving censorship accuracy. In the dataset of 1,000 chats in Minecraft- Server-Chat dataset, 108 leetspeak inputs were analyzed. The Enhanced Algorithm demonstrates an 82.4% censorship success rate for leetspeak-masked inputs, reducing the error rate to 2.8% compared to the Existing Algorithm’s 28.7%. This gives the enhanced algorithm the accuracy rate of 97.2% in filtering leetspeak-masked profanity compared to the 71.3% of the existing algorithm. These advancements made the Enhanced Algorithm a robust, context-aware and accurate in leetspeaks for moderating Minecraft chats, fostering a safer and more inclusive online environment.

Keyword : Algorithm, BERT transformer, Leetspeak detection, Profanity filtering, Toxicity scoring