Writer Profiling Without the Writer's Text

David Jurgens, Yulia Tsvetkov, Dan Jurafsky

Abstract: Social network users may wish to preserve their anonymity online by masking their identity and not using language associated with any particular demographics or personality. However, they have no control over the language in incoming communications. We show that linguistic cues in public comments directed at a user are sufficient for an accurate inference of that user’s gender, age, religion, diet, and even personality traits. Moreover, we show that directed communication is even more predictive of a user’s profile than the user’s own language. We then conduct a nuanced analysis of what types of social relationships are most predictive of users’ attributes, and propose new strategies on how individuals can modulate their online social relationships and incoming communications to preserve their anonymity.