3M: Multi-document Summarization Considering Main and Minor Relationship

Anonymous

Abstract: The multi-document summarization (MDS) is an important branch of information aggregation. Compared with the single-document summary (SDS), MDS has three major challenges: (1) MDS involves too large search space to capture the attention; (2) the input of MDS contains a lot of redundant information and more complex logical relationships; (3) the different opinions of documents bring contradictions. To complete these three main challenges, we combine the Transformer and the Maximal Marginal Relevance (MMR) to design Multi-document summarization considering Main and Minor relationship (3M) model. In this model, we take one document as the main body and use the information of other documents as an addition to modifying the generation of the summary. Therefore, we can reduce the search space and ignore the redundancy in the minor documents. Empirical results on the Multi-News and DUC 2004 dataset show that the 3M brings substantial improvements over several strong baselines, the manual evaluation shows that the generated abstract is fluent and can better express the content of the main document. In addition, by selecting different main documents, 3M can generate multiple abstracts with different styles for one set of documents.

Paper Type: long