For the task of classifying news articles into their subject type (finance, science, politics, etc), what is a current state-of-the-art approach (described in detail)? Describe the latency of this approach. Think of simpler, classical models that can achieve faster throughput while maintaining high accuracy for this task.