High compression rate text summarization

Abstract

This thesis focuses on methods for condensing large documents into highly concise summaries, achieving compression rates on par with human writers. While the need for such summaries in the current age of information overload is increasing, the desired compression rate has thus far been beyond the reach of automatic summarization systems. The potency of our summarization methods is due to their in-depth modelling of document content in a probabilistic framework. We explore two types of document representation that capture orthogonal aspects of text content. The first represents the semantic properties mentioned in a document in a hierarchical Bayesian model. This method is used to summarize thousands of consumer reviews by identifying the product properties mentioned by multiple reviewers. The second representation captures discourse properties, modelling the connections between different segments of a document. This discriminatively trained model is employed to generate tables of contents for books and lecture transcripts. The summarization methods presented here have been incorporated into large-scale practical systems that help users effectively access information online.