Compare predicted phrase boundaries and labels with the annotated corpus. This measures consistency with training data; accuracy on unseen sentences requires a separate human-reviewed test set.