Automatic Document Summarization via Deep Neural Networks

Automatic document summarization aim to extracting sentences which might cover the main content of a document or documents. To achieve this, many algorithms have been tried to rank the sentences by using task-specific features in a shallow architecture. The main challenge of these approaches is to keep balance between information coverage and redundancy because of absence of discovering the intrinsic semantic representation. Inspired by the recent successful achievement of Deep Learning, this paper proposes a new framework of document summarization via Deep Neural Networks (DNNs). Specifically, we feed the sentences as the input to the visible layer of DNNs. After pretraining layer by layer and fine-tuning, the lower dimensional semantic space can be revealed. Based on this space, we design sentences extraction algorithm to construct the summary. Experiments on the DUC2006 and DUC2007 dataset show that our framework works better than state-of-the-art methods.