Predicting Interpersonal Influence from Conversational Features

Interpersonal influence has a radical impact on the dissemination of information in online social media. Methods for measuring this influence between online conversation partners are often over-reliant on platform-level features, rendering them inoperable in other settings. We propose a novel and transferable solution using BERT to derive features of conversations that indicate influence. In an evaluation across a diverse discussion dataset, we show that our framework competes with existing large language models, being able to predict both social and behavioural measures of influence accurately, and at different levels of resolution, with a Macro-F1 above 0.91 in all cases of social influence.