Which metrics are commonly used to evaluate generation tasks in prompt-based systems?

• A. ROUGE and BLEU for generation, with accuracy and F1 for classification.
• B. Exponential loss and cross-entropy.
• C. Perplexity only.
• D. ROUGE and BLEU for generation, plus human evaluation.

Answer: (D)

Evaluating generated text focuses on how well the output matches content and how well it reads in practice. ROUGE and BLEU quantify overlap with reference texts, which is useful for tasks like summarization and translation, but they can miss nuances of fluency, coherence, and factual accuracy. Pairing those automatic metrics with human evaluation provides qualitative judgments on fluency, relevance, and correctness, giving a fuller picture of generation quality in prompt-based systems. Perplexity, while a helpful training measure of how predictably a model writes, doesn’t directly reflect the quality of generated outputs. Training losses like cross-entropy or exponential loss are optimization criteria, not evaluation of produced text. Classification metrics (accuracy, F1) apply to discriminative tasks, not generation. So combining ROUGE and BLEU with human evaluation best captures the strengths and weaknesses of generated prompts.