Metrics for natural language performance