In a confusion matrix for a multi-class NLP classifier, what do the diagonal entries represent?

• A. Correct predictions for each class.
• B. Misclassifications for each class.
• C. Predicted probabilities for each class.
• D. Ground truth counts for each class.

Answer: (A)

The diagonal entries show how many times the model predicted the exact same class as the true label for that class. In a multi-class confusion matrix, each diagonal cell counts the instances where the actual class and the predicted class match, i.e., the true positives for that class. These diagonal counts tell you how often the model is correct for each category, and they combine to give overall accuracy when summed and divided by the total number of samples.

Off-diagonal cells are misclassifications, where the true class is predicted as a different class. Predicted probabilities aren’t stored in the confusion matrix; they come from the model’s score outputs and can be used to compute other metrics, but the matrix itself uses counts (or normalized counts) of predictions. The total number of samples belonging to each class—the ground truth counts—is reflected in the row or column totals, not in the diagonal.