For algorithms, spotting trolling or harassment online largely relies on monitoring keywords -- a blunt instrument that doesn’t account for context or the evolution of conversations. But a new machine learning algorithm might change that.