[Paper Review] Conversations Gone Awry: Detecting Early Signs of Conversational Failure

Justine Zhang, Jonathan P. Chang+5 more|arXiv (Cornell University)|May 14, 2018

Hate Speech and Cyberbullying Detection41 references57 citations

TL;DR

This paper introduces the task of predicting early, from the first exchange, whether a civil online conversation will derail into personal attacks, using politeness and rhetorical prompts.

ABSTRACT

One of the main challenges online social systems face is the prevalence of antisocial behavior, such as harassment and personal attacks. In this work, we introduce the task of predicting from the very start of a conversation whether it will get out of hand. As opposed to detecting undesirable behavior after the fact, this task aims to enable early, actionable prediction at a time when the conversation might still be salvaged. To this end, we develop a framework for capturing pragmatic devices---such as politeness strategies and rhetorical prompts---used to start a conversation, and analyze their relation to its future trajectory. Applying this framework in a controlled setting, we demonstrate the feasibility of detecting early warning signs of antisocial behavior in online discussions.

Motivation & Objective

• Motivate the problem of predicting future antisocial behavior before it occurs in online conversations.
• Construct a controlled, labeled dataset of Wikipedia talk-page conversations that start civilly but may derail.
• Develop a framework to extract pragmatic devices like politeness strategies and rhetorical prompts.
• Assess whether these early linguistic cues can predict the conversation’s trajectory and provide actionable insights.

Proposed method

• Identify conversations that start civilly using a toxicity classifier and crowdsourced filtering.
• Create a matched, controlled set pairing awry-turning and on-track conversations on the same talk page to control topical confounds.
• Extract politeness strategies via pattern matching on dependency parses following Brown and Levinson (1987).
• Discover domain-specific conversation prompts through unsupervised, dependency-parse–based representation and clustering.
• Evaluate the predictive power of pragmatic features using logistic regression in a leave-one-page-out cross-validation.

Experimental results

Research questions

• RQ1 Can early linguistic cues in the initial exchange predict whether a conversation will derail into a personal attack?
• RQ2 How do politeness strategies relate to the future trajectory of a conversation?
• RQ3 Do domain-specific prompts discovered unsupervisedly contribute predictive signal beyond predefined politeness markers?
• RQ4 What is the relative predictive power of pragmatic features compared to baseline language or user characteristics?

Key findings

• Politeness and prompt-type features from the first two comments yield measurable predictive signal about future derailment.
• First-exchange direct questions and sentence-initial you pronouns associate with awry-turning conversations (p < 0.001).
• Gratitude and greetings in the first exchange are more common in on-track conversations (p < 0.001 to p < 0.05).
• Pragmatic features (prompts + politeness) achieve 61.6% accuracy in predicting future outcomes, outperforming baselines like bag-of-words and sentiment alone.
• Combining toxicity signals with pragmatic features reaches 64.9% accuracy, close to human performance (72%), and improves on some baselines.
• Humans correctly predict future attacks 72% of the time, and the model matches humans on 80% of the cases where humans were correct.