A little goes a long way: Improving toxic language classification despite data scarcity