Multi-doc Hybrid Summarization via Salient Representation Learning

Multi-document summarization is gaining more and more attention recently and serves as an invaluable tool to obtain key facts among a large information pool. In this paper, we proposed a multi-doc hybrid summarization approach, which simultaneously generates a human-readable summary and extracts corresponding key evidence giving a multi-doc input. To fulfill that purpose, we crafted a salient representation learning method to induce latent noteworthy features, which are effective for joint evidence extraction and summary generation. In order to train that model, we performed multi-task learning to optimize a composite loss, which is hierarchically constructed over the extractive and abstractive sub-components. We implemented such a fine system based on a ubiquitously-adopted transformer architecture and conducted experiments on a variety of datasets across two domains, achieving superior performance than the baselines.