At present, most of the ways for online platforms such as forums to obtain hate speech come from user reports. It is impossible for human beings to keep staring at those endless negative remarks. At the beginning of this year, Google developed perspective software, which can quickly identify abusive comments and facilitate manual review. Its working principle is based on the similarity between online comments and comments labeled as "toxic". But then the immaturity of software technology began to appear, and there were many disadvantages in the scoring mechanism. For example, some remarks are "As a girl, you are so smart", and the similarity with malicious remarks reaches18%; "I like Hitler" is only 2%. Say no to cyber violence, AI makes hate speech have nowhere to hide.
Different from this method based on keyword tags, the system developed by Canadian researchers takes a different approach. The system mainly studies speeches for African-Americans, obese people and women. On Reddit or Voat (a website similar to Reddit), comments on these people abound. The team found two most active communities: one likes to make bad comments and the other likes to make friendly comments. They use artificial intelligence software to learn the phonetic features of the members of these two communities and improve the system's ability to correctly identify negative remarks.
The research results show that this method is more accurate than the system based on keyword tagging, and there is almost no misjudgment. Some speeches do not contain conventional insulting words, but they are also hate speeches. If it was unrecognizable by the previous method, it can be done now. For example, "I don't think there is anything wrong with this. Animals always attack each other. " This sentence is systematically marked as hate speech, because the word "animal" here means racial insult.