This paper proposes a process of modelling interpersonal influence in discourse that leverages spatial relationships inherent in term- and document-embedding, while focusing on identifying group leaders. The proposed method transforms speakers' comments from an embedding space, which is speaker-and time-agnostic, to a normalized influence space. Pilot experiments on two group discussions assessed modelling stability under perturbations and tested our algorithms. The results highlight the merits of the method in identifying the most influential participants, i.e., the “ringleaders”.