Planning with Learned Entity Prompts for Abstractive Summarization

Abstract We introduce a simple but flexible mechanism to learn an intermediate plan ground the generation of abstractive summaries. Specifically, we prepend (or prompt) target summaries with entity chains—ordered sequences entities mentioned in summary. Transformer-based sequence-to-sequence models are then trained generate chain and continue generating summary conditioned on input. experimented both pretraining finetuning this content planning objective. When evaluated CNN/DailyMail, XSum, SAMSum, BillSum, demonstrate empirically that grounded objective improves specificity for all datasets, achieves state-of-the-art performance XSum SAMSum terms rouge. Moreover, chains provides control hallucinations By prompting decoder modified drops hallucinated entities, outperform approaches faithfulness when automatically by humans.