Alphabet company Jigsaw and the Wikimedia Foundation teamed up last year to study online harassment in effort to develop an automated system to detect it. Now they’re opening up their collection of over 100,000 labeled comments to help researchers train machine-learning algorithms to know what abuse looks like. An additional 95 million article and talk comments from 2001-2015 were also released.
Crowdsourced human editors determined that roughly 11% of the more than 100,000 comments contained personal attacks, according to their recently published study. That data was then used to train algorithms. “With testing, we found that a fully trained model achieves better performance in predicting whether an edit is a personal attack than the combined average of 3 human crowd-workers,” states a release from researchers at both Jigsaw and Wikimedia.
Though exciting, researchers admit this is a baby shoe-sized step forward—the research only concerned obvious instances of abuse.