Data Augmentation for Abstractive Query-Focused Multi-Document Summarization

The progress in Query-focused Multi-Document Summarization (QMDS) has been limited by the lack of sufficient largescale high-quality training datasets. We present two QMDS datasets, which we construct using data augmentation methods: (1) transferring commonly used single-document CNN/Daily Mail summarization dataset to create QMDSCNN dataset, and (2) mining search-query logs QMDSIR dataset. These datasets have complementary properties, i.e., real summaries but queries are simulated, while simulated summaries. To cover both these summary query aspects, build abstractive end-to-end neural network models on combined that yield new state-of-the-art transfer results DUC also introduce hierarchical encoders enable a more efficient encoding together with multiple documents. Empirical demonstrate our methods outperform baseline automatic metrics, as well human evaluations along attributes.