Conversation Modeling to Predict Derailment

Abstract

Conversations among online users sometimes derail, i.e., break down into personal attacks. Derailment interferes with the healthy growth of communities in cyberspace. The ability to predict whether an ongoing conversation will derail could provide valuable advance, even real-time, insight to both interlocutors and moderators. Prior approaches predict conversation derailment retrospectively without the ability to forestall the derailment proactively. Some existing works attempt to make dynamic predictions as the conversation develops, but fail to incorporate multisource information, such as conversational structure and distance to derailment. We propose a hierarchical transformer-based framework that combines utterance-level and conversation-level information to capture fine-grained contextual semantics. We propose a domain-adaptive pretraining objective to unite conversational structure information and a multitask learning scheme to leverage the distance from each utterance to derailment. An evaluation of our framework on two conversation derailment datasets shows an improvement in F1 score for the prediction of derailment. These results demonstrate the effectiveness of incorporating multisource information for predicting the derailment of a conversation.