Learning from Reduced Labels for Long-Tailed Data