Dynamic Forecasting of Conversation Derailment

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7915–7919 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7915 Dynamic Forecasting of Conversation Derailment Yova Kementchedjhieva University of Copenhagen yova@di.ku.dk Anders Søgaard University of Copenhagen soegaard@di.ku.dk Abstract Online conversations can sometimes take a turn for the worse, either due to systematic cultural differences, accidental misunderstand- ings, or mere malice. Automatically forecast- ing derailment in public online conversations provides an opportunity to take early action to moderate it. Previous work in this space is limited, and we extend it in several ways. We apply a pretrained language encoder to the task, which outperforms earlier approaches. We further experiment with shifting the train- ing paradigm for the task from a static to a dynamic one to increase the forecast horizon. This approach shows mixed results: in a high- quality data setting, a longer average forecast horizon can be achieved at the cost of a small drop in F1; in a low-quality data setting, how- ever, dynamic training propagates the noise and is highly detrimental to performance. 1 Introduction The flow of a conversation can be interrupted when its participants turn to attacking each other as op- posed to communicating constructively. The task of forecasting derailment aims to predict outbursts of hostile behavior between users in online con- versations before they happen (Zhang et al., 2018). This stands in opposition to related tasks like hate speech detection (Fortuna et al., 2020), where abu- sive language is identified post factum. Accurate forecasts could give forum moderators an early warning of troublesome exchanges or even serve as a step towards the prevention, human-driven or automatic, of derailment. Table 1 shows a conver- sation that starts off civil and then derails into a personal attack. As a personal attack can occur at any point dur- ing an unfolding conversation, forecasting has to happen dynamically. Furthermore, the earlier the warning for a potential derailment comes, the more actionable it is. In the example from Table 1, we would therefore attempt a prediction already after A Where’s your source? B I’m just looking for it. A In other words, it’s all bullshit you pulled outta your ass, right? Table 1: A conversation derails into a personal attack. the first comment made by user A, and hope to get an early warning for the derailment coming up in two turns. This is indeed the inference procedure that was used in previous work on the task (Zhang et al., 2018; Chang and Danescu-Niculescu-Mizil, 2019). Training, on the other hand, was done stat- ically in those works, i.e. by feeding all turns up to the personal attack as input and making a pre- diction based on that. This leads to a disparity be- tween the training and inference procedures, where at training time the implicit task of the model is to predict whether the next turn contains a personal attack, whereas at inference time its task is to pre- dict whether any future turn will contain a …

Content truncated to 500 words