Temporal Multi-Document Summarization Based on Macro-Micro Importance Discriminative Model

Temporal multi-document summarization (TMDS) aims to capture the evolving information of relevant document sets across periods. Different from the traditional static multi-document summarization, it handles the dynamical collection relevant to a topic. How to resolve the key problems in the temporal context is a new challenge. This paper focuses on how to summarize the series news reports by a generic and extractive way. According to the temporal characteristics of series news reports at different levels of topical detail, a content selection method based on the macro-micro importance discriminative model is proposed. This method mines the temporal characteristics of series news reports from macro and micro views in order to provide the cue for content selection. Firstly, important time points are selected based on the macro importance discriminative model; then important sentences are selected by the micro importance discriminative model; and then these two models are integrated into a macro-micro importance discriminative model. Lastly, summary sentences are ordered chronologically. The experimental results on five groups of Chinese news corpus prove that this method is effective. It also shows that the macro and micro temporal characteristics of series news have the recursive property to some extent and macro coarse filtering helps to the content selection of TMDS.