Fast document summarization using locality sensitive hashing and memory access efficient node ranking

Abstract

Text modeling and sentence selection are the fundamental steps of a typical extractive document summarization algorithm. The common text modeling method connects a pair of sentences based on their similarities. Even thought it can effectively represent the sentence similarity graph of given document(s) its big drawback is a large time complexity of $O(n^2)$, where n represents the number of sentences. The quadratic time complexity makes it impractical for large documents. In this paper we propose the fast approximation algorithms for the text modeling and the sentence selection. Our text modeling algorithm reduces the time complexity to near-linear time by rapidly finding the most similar sentences to form the sentences similarity graph. In doing so we utilized Locality-Sensitive Hashing, a fast algorithm for the approximate nearest neighbor search. For the sentence selection step we propose a simple memory-access-efficient node ranking method based on the idea of scanning sequentially only the neighborhood arrays. Experimentally, we show that sacrificing a rather small percentage of recall and precision in the quality of the produced summary can reduce the quadratic to sub-linear time complexity. We see the big potential of proposed method in text summarization for mobile devices and big text data summarization for internet of things on cloud. In our experiments, beside evaluating the presented method on the standard general and query multi-document summarization tasks, we also tested it on few alternative summarization tasks including general and query, timeline, and comparative summarization.