Virtual moderator using analysis of user trajectories on a semantic map based on convolutional neural networks

Comments in social media require moderation to avoid escalation of insults and overt aggression. Given the current large volume of communications, this moderation can only be done automatically. Modern state-of-the-art techniques that allow one to automatically detect insults in text have their limitations. For example, methods using forbidden words may not be efficient in online chats. The main drawback, however, is related to the fact that evaluation is done without taking into account user’s behavior in the past. The main purpose of the present work is to develop a virtual moderator that is free of this flaw. The approach is based on a set of convolutional neural networks and a semantic map, together taken as the basis of the model. The networks are trained for sentiment extraction from sentences found in available corpora. The virtual moderator matches coordinates of the 2-D semantic map to each fragment of the given text. One map axis represents offensiveness of the fragment, and the other the sentiment (positivity or negativity). The virtual moderator takes into account behavioral history of each user. However, newer messages determine the output to a higher extent compared to the older ones. The implemented prototype of a virtual moderator was tested on conversations in the Telegram app. Participants were instructed to communicate freely. At the same time, a confederate was introduced into the group, whose role was to play a bully, to insult others, and otherwise behave inappropriately. Afterwards, regular participants were asked to evaluate the correctness of Virtual Moderator’s decisions. The results show 73% accuracy of the Virtual Moderator.