The Power of Summary-Source Alignments

Multi-document summarization (MDS) is a challenging task, often decomposed tosubtasks of salience and redundancy detection, followed by text generation. Inthis context, alignment of corresponding sentences between a reference summaryand its source documents has been leveraged to generate training data for someof the component tasks. Yet, this enabling alignment step has usually beenapplied heuristically on the sentence level on a limited number of subtasks. Inthis paper, we propose extending the summary-source alignment framework by (1)applying it at the more fine-grained proposition span level, (2) annotatingalignment manually in a multi-document setup, and (3) revealing the greatpotential of summary-source alignments to yield several datasets for at leastsix different tasks. Specifically, for each of the tasks, we release a manuallyannotated test set that was derived automatically from the alignmentannotation. We also release development and train sets in the same way, butfrom automatically derived alignments. Using the datasets, each task isdemonstrated with baseline models and corresponding evaluation metrics to spurfuture research on this broad challenge.