Which statement best describes the bias-variance tradeoff and its impact on model generalization in NLP?

• A. High bias leads to underfitting; high variance leads to overfitting.
• B. High bias leads to overfitting; high variance leads to underfitting.
• C. Balanced bias and variance causes overfitting.
• D. Neither bias nor variance affects generalization.

Answer: (A)

In this question, the idea is how model complexity interacts with data to affect performance on new text data. Bias is the error from making strong assumptions about the data; high bias means the model is too simple to capture the true linguistic patterns, leading to underfitting where both training and unseen data are predicted poorly. Variance is the error from being overly sensitive to the training data; high variance means the model ends up fitting noise in the training set, resulting in overfitting and poor generalization to new text.

In NLP, you want enough capacity to understand syntax, semantics, and context, but not so much freedom that the model just memorizes the training corpus. The statement that best describes the tradeoff is that high bias causes underfitting and high variance causes overfitting, which directly ties to how well the model generalizes to unseen data.

The alternative interpretations are off: swapping the effects (bias causing overfitting or variance causing underfitting) misstates the relationships, suggesting the opposite behavior. Saying that a balanced bias and variance causes overfitting is incorrect because balancing often improves generalization, not overfitting. Finally, claiming neither bias nor variance affects generalization ignores a fundamental aspect of learning and model evaluation.