REDIRE: Extreme REduction DImension for extRactivE Summarization

Abstract

This paper presents an automatic unsupervised summarization model capable of extracting the most important sentences from a corpus. The unsupervised aspect makes it possible to do away with large corpora, made up of documents and their reference summaries, and to directly process documents potentially made up of several thousand words. To extract sentences in a summary, we use pre-entrained word embeddings to represent the documents. From this thick cloud of word vectors, we apply an extreme dimension reduction to identify important words, which we group by proximity. Sentences are extracted using linear constraint solving to maximize the information present in the summary. We evaluate the approach on large documents and present very encouraging initial results.