Considering that many online discussions have the potential to become contentious, researchers from Cornell University and the Wikimedia Foundation have created a template of sorts to predict when an online discussion might deteriorate.

Publishing the details of their research in arXiv, the team noted that many online conversations can erupt into arguments and, ultimately, personal attacks. This is particularly true on sites such as Wikipedia where editors, amateur and otherwise, offer critiques of work submitted by others in an effort to improve content on the site. Yet, many of the authors don’t take the critiques well and will, oftentimes, respond by posting negative comments.

The Wikimedia Foundation would like to keep such conversations constructive, steering comments away from the negative, not only for the benefit of those engaged in the argument but to also keep the site from developing a bad reputation. As such, they have teamed with researchers from Cornell to develop a computer system that can recognize when a conversation veers into negative territory, either redirecting it or halting the entire conversation.

Looking at more than 1,000 online discussions on the Wikipedia Talk pages, the system has been programmed to look for “cues” of polite discussion, focusing on words such as “please” and “thanks,” suggesting that their use would not likely result in the conversation spiraling into negative territory.

Conversely, discussions that began with direct questions or the word “you” had a much better chance of degrading the conversation as they are often associated with being contentious and hostile.

Applying that data to an algorithm, the system was able to detect, early in the conversation, when a conversation was apt to become negative with a rate of accuracy of 61.6 percent.