A friend thinks you can answer this question about AI Summarization Quality
Your meeting summarization product reports a ROUGE-L score of 0.42 and customer complaints about hallucinated attendee names and made-up action items. The team proposes to fine-tune for higher ROUGE. What's the right diagnosis and fix?