Personalized PageRank Based Multi-document Summarization

This paper presents a novel multi-document summa-rization approach based on Personalized PageRank (PPRSum). In this algorithm, we uniformly integrate various kinds of information in the corpus. At first, we train a salience model of sentence global features based on Naïve Bayes Model. Secondly, we generate a relev-ance model for each corpus utilizing the query of it. Then, we compute the personalized prior probability for each sentence in the corpus utilizing the salience model and the relevance model both. With the help of personalized prior probability, a Personalized PageRank ranking process is performed depending on the relationships among all sentences in the corpus. Additionally, the redundancy penalty is imposed on each sentence. The summary is produced by choosing the sentences with both high query-focused information richness and high information novelty. Experiments on DUC2007 are performed and the ROUGE evaluation results show that PPRSum ranks between the 1st and the 2nd systems on DUC2007 main task.