[Spartacus]

Troll detection through emotional persistence in IRC chats

Schema of an IRC networkThe Internet Relay Chat protocol has been one of the most important methods for real-time communication since the beginning of the Internet. Although not as popular as Social Networking Sites, the current IRC networks serve more than 500.000 users at a time (irc.netsplit.de, 2011).  In the 90′s, IRC channels were used as a way to avoid media blackouts during the Soviet Union coup and the Gulf war. We tend to think that these “Twitter revolutions” are a new tendency of social networks, but distributed anonymous online communication in political events  existed in IRC chats way before the Web 2.0. Normal users were able to communicate with each other safely away from traditional media, although these chats lacked the reach and pervasiveness provided by nowadays smartphones and social networks.  Currently, IRC channels make it to the news quite often, as they are one of the main channels of communication of Internet activism groups like Anonymous. Due to their strong anonymity, the purposes for the use of these chats range from military applications to dating.

Schema of emotional expression in IRC chats In the article “Emotional persistence in online chatting communities” (Nature Scientific Reports, 2012),  we study the role of anonymity in IRC channels. We look into the different patterns of emotional expression of the general chat discussions, looking for what are the features of human communication under the anonymity of IRC. Our dataset contains the general discussion messages of 42 days in 20 different IRC channels, with more than 2.5 million messages created by more than 20.000 users. Each one of these messages has been processed with the state-of-the-art emotion detection technique SentiStrength, extracting the emotional content of the message as positive (+), negative (-), or neutral(0). This way, the chat discussion is a stream of emotional values in time that can be quantitatively analyzed.

We found that the time difference between the messages of each user follows universal rules that are present in many other means of communication, from sms exchanges to the letter correspondence of Darwin and Einstein. In addition, statistical analysis of the time between messages in the general discussion shows the existence of memory in the community, indicating that users remember each other. We analyzed the emotional expression of each user, extracting a measure called persistence. If the persistence of a user is above 0.5, the user tends to follow previous emotional expressions, reporting emotional episodes to the other users in the community. If the persistence is below 0.5, the user does the opposite, so the next messages will tend to have different emotional content, balancing this way the overall emotions contained in her messagesPersistence versus average emotional expression of users

Most of the chat users express positive emotions, as the mean polarity of their messages is above 0. In addition, most of the users are also persistent. Our method allows the detection of users that consistently express negative emotions, with a mean polarity below 0, and show persistence in their emotional content. This way, we have invented a Troll detector, that tells when a user behaves extremely negatively and not following the usual social norms that tend to positivity. In particular, 6.1 % of the users included in our study showed this negative pattern. If this technique would be applied to private conversations, we could detect users with depressive tendencies as they would consistently express negative emotions to their closely related peers.

Furthermore, we designed an Agent-based model that reproduces the emergence of emotional persistence in the whole chat discussion. This general persistence is a trace of the existence of collective emotions, which are emotional states shared by a large amount of the users of the chatroom due to the way they interact. This model is currently being used by our partners in OFAI to design the next generation of dialog systems (chatbots), which will include modules of emotional intelligence that imitate human behavior when taking part of a general online discussion.

http://dgarcia.eu/?p=12

tl:dr version. They designed a model to detect trolls in IRC using persistence as a measure. Trollplayer your trolling days are over! ;D