What is the primary bottleneck in text classification?