lol that's pretty much it. In that case you'd need to get better labeled data (considering supervised learning).