Document Summarization with Latent Queries

Abstract The availability of large-scale datasets has driven the development neural models that create generic summaries for single or multiple documents. For query-focused summarization (QFS), labeled training data in form queries, documents, and is not readily available. We provide a unified modeling framework any kind summarization, under assumption all are response to query, which observed case QFS latent summarization. model queries as discrete variables over document tokens, learn representations compatible with unobserved query verbalizations. Our formulates generative process, jointly optimizes conditional language model. Despite learning from only, our approach outperforms strong comparison systems across benchmarks, types, settings, target domains.1