Conversation Derailment Forecasting with Graph Convolutional Networks

Online conversations are particularly susceptible to derailment, which can manifest itself in the form of toxic communication patterns like disrespectful comments or verbal abuse. Forecasting conversation derailment predicts signs of derailment in advance enabling proactive moderation of conversations. Current state-ofthe-art approaches to address this problem rely on sequence models that treat dialogues as text streams. We propose a novel model based on a graph convolutional neural network that considers dialogue user dynamics and the influence of public perception on conversation utterances. Through empirical evaluation, we show that our model effectively captures conversation dynamics and outperforms the state-of-the-art models on the CGA and CMV benchmark datasets by 1.5% and 1.7%, respectively.

Introduction. The widespread availability of chat or messaging platforms, social media, forums and other online communities has led to an increase in the number of online conversations between individuals and groups. In contrast to offline or face-to-face communication, online conversations require moderation to maintain the integrity of the platform and protect users’ privacy and safety (Kilvington, 2021). Moderation can help to prevent harassment, trolling, hate speech, and other forms of abusive behavior (Tontodimamma et al., 2021). It can also help to prevent and address conversation derailment. Conversation derailment refers to the process by which a conversation or discussion is redirected away from its original topic or purpose, typically as a result of inappropriate or off-topic comments or actions by one or more participants. In online conversations, derailment can be exacerbated by the lack of nonverbal cues and the perceived anonymity that can be provided by the internet.

Discussion / Conclusion. Unlike previous models which were based on simpler sequence models, FGCN is built on a graph convolutional neural network and is able to capture the dynamics of multi-party dialogue, including user relationships and public perception of conversation statements. FGCN performed significantly better than state-of-the-art models on two widely used benchmark datasets, CGA and CMV. Conversation derailment is a significant issue that frequently and severely impacts our online social interactions, whether in casual settings or more formal contexts such as online learning or remote work. The ability to accurately predict derailment has the potential to enhance the effectiveness of moderation and thus protect individuals who are vulnerable to emotional abuse or harm and improve the overall quality of online interactions. Graph models require four or more utterances to form meaningful conversation connections and model their dynamics. In some cases, conversations that derail are not sufficiently long and may be best modeled by simpler sequential models. Any of these models will work best with asynchronous conversations where there is a time lag between the turns to allow for moderation after forecasting. In our paper, we focus on the problem of forecasting conversation derailment.