SCHC: Incorporating Social Contagion and Hashtag Consistency for Topic-Oriented Social Summarization

The boom of social media platforms like Twitter brings the large scale short, noisy and redundant messages, making it difficult for people to obtain essential information. We study extractive topic-oriented social summarization to help people grasp the core information on social media quickly. Previous methods mainly extract salient content based on textual information and shallow social signals. They ignore that user generated messages propagate along the social network and affect users on their dissemination path, leading to user-level redundancy. Besides, hashtags on social media are a special kind of social signals, which can be regarded as keywords of a post and contain abundant semantics. In this paper, we propose to leverage social theories and social signals (i.e. multi-order social relations and hashtags) to address the redundancy problem and extract diverse summaries. Specifically, we propose a novel unsupervised social summarization framework which considers Social Contagion and Hashtag Consistency (SCHC) theories. To model relations among tweets, two relation graphs are constructed based on user-level and hashtag-level interaction among tweets. These social relations are further integrated into a sparse reconstruction framework to alleviate the user-level and hashtag-level redundancy respectively. Experimental results on the CTS dataset prove that our approach is effective.