Continuous Summarization over Microblog Threads

Abstract

With the dramatic growth of social media users, microblogs are created and shared at an unprecedented rate. The high velocity and large volumes of short text posts (microblogs) bring redundancies and noise, making it hard for users and analysts to elicit useful information. In this paper, we formalize the problem from a summarization angle – Continuous Summarization over Microblog Threads (CSMT), which considers three facets: information gain of the microblog dialogue, diversity, and temporal information. This summarization problem is different from the classic ones in two aspects: (i) It is considered over a large-scale, dynamic data with high updating frequency; (ii) the context between microblogs are taken into account. We first prove that the CSMT problem is NP-hard. Then we propose a greedy algorithm with ($1-1/\mathrm{e}$) performance guarantee. Finally we extend the greedy algorithm on the sliding window to continuously summarize microblogs for threads. Our experimental results on large-scale datasets show that our method is more superior than other two baselines in terms of summary diversity and information gain, with a close time cost to the best performed baseline.