Analysis of Factors Contributing to Conversational Derailment Using Text Style Embeddings

This study aims to comprehend the phenomenon in which conversations on online platforms deviate from the original conversational theme and evolve into personal attacks. The central focus is the effect of the textual style of the utterances in a conversation on the conversational derailment. To this end, we have developed a logistic regression model that classifies conversational derailment using the principal components of interpretable text-style embeddings. Consequently, our study demonstrated that the conversational style that effectively prevents derailment is characterized by logical support of ideas and acquiring information from conversation partners through expressions of curiosity. This conversation style involves using questions to explore the other person’s thoughts and background rather than just imposing one’s own opinions. It respects the conversation partner’s perspective and deepens the dialogue. These findings enabled platform moderators to define their moderation policy more effectively when detecting conversations that exhibited potential for derailment.