Open Domain Multi-document Summarization: A Comprehensive Study of Model Brittleness under Retrieval

Multi-document summarization (MDS) assumes a set of topic-related documents are provided as input. In practice, this document set is not always available; it would need to be retrieved given an information need, i.e. a question or topic statement, a setting we dub "opendomain" MDS. We study this more challenging setting by formalizing the task and bootstrapping it using existing datasets, retrievers and summarizers. Via extensive automatic and human evaluation, we determine: (1) state-of-theart summarizers suffer large reductions in performance when applied to open-domain MDS, (2) additional training in the open-domain setting can reduce this sensitivity to imperfect retrieval, and (3) summarizers are insensitive to the retrieval of duplicate documents and the order of retrieved documents, but highly sensitive to other errors, like the retrieval of irrelevant documents. Based on our results, we provide practical guidelines to enable future work on open-domain MDS, e.g. how to choose the number of retrieved documents to summarize. Our results suggest that new retrieval and summarization methods and annotated resources for training and evaluation are necessary for further progress in the open-domain setting. 1 How to assemble the document index? For our purposes, we take the set of all documents in the train, validation, and test splits of each dataset to form D index . This guarantees that the ground-truth documents for each example are present in the index while providing plenty of negative examples.

How many documents to summarize? The number of retrieved documents to summarize, k, 6 See the BEIR (Thakur et al., 2021) zero-shot benchmark 7 Except for MSˆ2, where we found the provided "background" section to perform better as a query; see Appendix A 8 We also experimented with query generation using LLMs (e.g. GPT-3), but found that they significantly underperformed the reference summary as query, e.g. by at least 8 points P/R@K on a sample of the Multi-News validation set