Microblog summarization using Paragraph Vector and semantic structure

Abstract

There are two fundamental difficulties that are still hindering the development of microblog summarization. The first problem is the features sparseness of microblog, which restricts the performance of sub-topics detection. The second one is the sentence selection from sub-topics that is based mainly on centrality approaches to measure sentence salience. Also, the semantic features and relations features between sentences and sub-topics were not given much attention. In order to address the two aforementioned problems, we propose a summarization method considering Paragraph Vector and semantic structure. Firstly, we construct sentence similarity matrix that involves the contextual information of microblogs to detect sub-topics by using Paragraph Vector. Secondly, we analyze the sentences by utilizing Chinese Sentential Semantic Model (CSM) to get semantic features; then the relations features are obtained based on the similarity matrix and semantic features above. Finally, the most informative sentences can be selected accurately from microblogs belonging to the same sub-topics by semantic features and relation features. The experimental results show that the ROUGE-1 value is up to 53.17% with 1.5% compression ratio. The results indicate that applying Paragraph Vector to the field of microblog summarization can effectively improve sub-topics detection. Additionally, semantic features and relation features enhance summarization result jointly. Furthermore, CSM provides a promising tool for sentence semantic analysis.