Unsupervised personality recognition for social network sites

In this paper, we present a system for personality recognition that exploits linguistic cues and does not require supervision for evaluation. We run the system on a dataset sampled from a popular Social Network: FriendFeed. We adopted the five classes from the standard model known in psychology as the ``Big Five'': extraversion, emotional stability, agreeableness, conscientiousness and openness to experience. Making use of the linguistic features associated with those classes the system generates one personality model for each user. The system then evaluates the models by comparing all the posts of one single user (users that have only one post are discarded). As evaluation measures the system provides accuracy (measure of the reliability of the personality model) and validity (measure of the variability of writing style of a user). The analysis of a sample of 748 Italian users of FriendFeed showed that that the most frequent personality type is represented by the model of an extravert, insecure, agreeable, organized and unimaginative person.