Estimating Personality from Social Media Posts

Abstract

An individual's personality determines the probable repertoire of their reactions to a particular situation. A social robot is much more effective if it is able to learn and so take into account the properties of the humans around it, including personalities. We investigate how well personality can be estimated based on modest amounts of speech or writing, which a social robot might (over)hear. Such a technique also permits humans to be able to infer the personalities of other humans 'at a distance' based on their writing in political, hiring, negotiation, and other relationship settings. We design and implement a technique for predicting personality from small amounts of text, with accuracies comparable to inter-human agreement and substantially better than previous algorithmic approaches (except for a few that use much richer data). The technique works for both of the popular personality typologies, the Big Five and the Myers-Briggs. Because the approach does not require a lexicon, it is language independent. We illustrate using eight different languages, including Arabic.