Exploiting Sparse Topics Mining for Temporal Event Summarization

Abstract

Information explosion, both in cyberspace world and real world, nowadays has brought about pressing needs for comprehensive summary of information. The challenge for constructing a quality one lies in filtering out information of low relevance and mining out highly sparse relevant topics in the vast sea of data. It is a typical imbalanced learning task and we need to achieve a precise summary of temporal event via an accurate description and definition of the useful information and redundant information. In response to such challenge, we introduced: (1) a uniform framework of temporal event summarization with minimal residual optimization matrix factorization as its key part; and (2) a novel neighborhood preserving semantic measure (NPS) to capture the sparse candidate topics under that low-rank matrix factorization model. To evaluate the effectiveness of the proposed solution, a series of experiments are conducted on an annotated KBA corpus. The results of these experiments show that the solution proposed in this study can improve the quality of temporal summarization as compared with the established baselines.