Feature-based Unsupervised Method for Salient Sentence Ranking in Text Summarization Task

Salient Sentence Ranking is an essential task that plays a vital role in Data Mining, especially in unsupervised document summarization tasks. In this paper, we introduce a simple yet effective unsupervised method to extract the salient sentences from a cluster of documents. Our method synthesizes the sentence scoring from various feature-based information containing position, topic, keyword, semantic, entity, sentence centroid -scores. The proposed method has the potential to generate large-scale pseudo-summary, which supports the tasks of summarization. To this end, our approach is able to incorporate pre-trained objectives used in pre-trained language models to diminish the problems of the lack of annotated datasets in low-resource languages like Vietnamese. We also conducted experiments to verify the effectiveness of various feature-based scoring methods and their combinations. Our experimental results on two well-known benchmark datasets, MultiNews and NewSHead, show the superiority of our proposed method compared with the previous unsupervised approaches.