ROUGE (e.g., ROUGE-L) is the correct choice because it prioritizes recall and measures overlap with reference summaries, which helps confirm that generated summaries include the key customer concerns.

ROUGE variants compute n gram overlap and ROUGE L uses longest common subsequence to capture coverage and ordering. This recall orientation makes ROUGE well suited when the evaluation goal is to ensure that important points from the full customer responses are preserved in a brief summary.

BLEU metric is not ideal because it is precision oriented and it was developed for machine translation. It rewards exact matches and can penalize valid paraphrases, so it does not align with a recall focused coverage requirement.

Embedding-based similarity can indicate overall semantic closeness between texts but it does not directly measure whether specific reference points are present. That limitation makes it less reliable for confirming coverage of key information.

Token-to-summary efficiency measures brevity and cost and it does not assess content coverage. It can help control length but it cannot confirm that the summary preserves all important customer concerns.

Full AWS Practitioner Certification Question