Conversations Gone Awry: Detecting Early Signs of Conversational Failure

Justine Zhang, Jonathan P. Chang, Cristian Danescu-Niculescu-Mizil, Lucas Dixon, Yiqing Hua, Nithum Thain, Dario Taraborelli

Proceedings of ACL 2018.


Data and Code

Fun: Guess whether a conversation will go awry

Talk slides and video: presented by Justine Zhang at ACL 2018

Talk video: presented by Jonathan P. Chang at the Wikimedia Research Showcase                                   

Media coverage:

                                   The Verge: Machine learning is helping computers spot arguments online before they happen

                                   MIT Technology Review: Machine learning could stop an online war of words before it starts

                                   Fast Company: Scientists are building a detector for conversations likely to go bad

Related research:         

                                    Anti-Social Computing

                                    Conversational Behavior



Which of the following conversations between Wikipedia editors will end with a personal attack?


A1: Why there's no mention of it here? Namely, an altercation with a foreign intelligence group? True, by the standards of sources some require it wouln't even come close, not to mention having some really weak points, but it doesn't mean that it doesn't exist.

    A2: So what you're saying is we should put a bad source in the article because it exists?


B1: Is the St.Petersberg Times considered a reliable source by wikipedia? It seems that the bulk of this article is coming from that one article, which speculates about missile launches and UFOs. I'm going to go through and try and find corroborating sources and maybe do a rewrite of the article. I don't think this article should rely on one so-so source.

    B2: I would assume that it's as reliable as any other mainstream news source.


Hint: the attack is “Wow, you're coming off as a total d**k. [...]  What the hell is wrong with you?''



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.




  author={Justine Zhang, Jonathan P. Chang, Cristian Danescu-Niculescu-Mizil,

                    Lucas Dixon, Yiqing Hua, Nithum Thain, Dario Taraborelli},

  title={Conversations Gone Awry: {Detecting} Early Signs of Conversational Failure},

  booktitle={Proceedings of ACL},