Tracing Social Influence in Online Hate-Speech Discussions

A Hybrid Agent-Based Modeling and NLP Approach

Hate-speech dialogues are emotionally charged social interactions characterized by hostility, anger, and strong dislike between participants. Under such conditions, individuals may not assimilate toward the opinions they encounter. Instead, dislike of another participant may cause them to reject that participant’s position and move further away through negative influence.

In this project, we investigate whether dislike-induced negative influence occurs more frequently in hate-speech dialogues and whether it contributes to escalating polarization. We combine several Natural Language Processing (NLP) methods: machine-learning-based text classification to estimate the degree of transphobia or anti-immigrant attitudes expressed in individual online posts, topic modeling to identify the thematic structure and topic shifts within dialogues, and sentiment and emotion analysis to capture negativity, anger, and hostility between participants. These measures are linked to reply relations, authorship, and timestamps to reconstruct how participants’ expressed attitudes change during interactions.

Building on our recent JASSS study of negative influence, we use the resulting longitudinal data to develop and evaluate agent-based models of online social influence. By comparing mechanisms such as assimilation, reinforcement, and negative influence, the project examines whether the distinctive emotional conditions of hate-speech dialogues make participants more likely to move apart.

Relevant Research: