A new machine-learning system tries to predict whether an online conversation is going to get nasty right from the get-go.
How it works: Researchers gathered more than 1,200 exchanges from the discussion sections of Wikipedia Talk pages. They went through and labeled different “linguistic cues” in the conversations, including attempts at politeness, like using “please” and “thanks,” or other phrases suggesting that debate was welcome, like “I believe” or “I think.” Using the tagged threads, they then trained a system to predict from the first comment if a conversation was going to go south.
Results: Humans were successful about 72 percent of the time at the task, compared with 61.6 percent for the algorithm. Not great, but the work uncovered some trends. For example, comments that have direct questions or start with the word “you” are signals that the conversation will end up getting heated.
Why it matters: An AI that predicts a conversation’s trajectory could help companies (cough, Twitter, cough) build tools that stop a fight or salvage online dialogue.
