Modeling Influence in Online Multi-party Discourse

In this article, we present our novel approach towards the detection and modeling of complex social phenomena in multi-party discourse, including leadership, influence, pursuit of power and group cohesion. We have developed a two-tier approach that relies on observable and computable linguistic features of conversational text to make predictions about sociolinguistic behaviors such as Topic Control and Disagreement, that speakers deploy in order to achieve and maintain certain positions and roles in a group. These sociolinguistic behaviors are then used to infer higher-level social phenomena such as Influence, which is the focus of this paper. We show robust performance results by comparing our computational results to participants' own perceptions and rankings of influence. We use weights learnt from correlations with known influence rankings to compute and score sociolinguistic behaviors and show performance significantly above baseline for two data sets and two different languages.