Inclusion-Based and Exclusion-Based Approaches in Graph-Based Multiple News Summarization

摘要

As combination of information extraction and relation analysis, constructing a comprehensive summary from multiple documents is a challenging task. Towards summarization of multiple news articles related to a specific event, an ideal summary should include only important common descriptions of these articles, together with some dominant differences among them. This paper presents a graph-based summarization method which is composed of text preprocessing, text-portion segmentation, weight assignment of text portions, and relation analysis among text portions, text-portion graph construction, and significant portion selection. In the process of portion selection, this paper proposes two alternative methods; inclusion-based and exclusion-based approach. To evaluate these approaches, a set of experiments are conducted on fifteen sets of Thai political news articles. Measured with ROUGE-N, the result shows that the inclusion-based approach outperforms the exclusion-based one with approximately 2% performance gap (80.59 to 78.21%).