Abstract  Graph-based manifold-ranking methods have been  successfully applied to topic-focused  multi-document summarization. This paper further  proposes to use the multi-modality manifold-ranking algorithm for extracting topic-focused  summary from multiple documents by considering  the within-document sentence relationships and the  cross-document sentence relationships as two  separate modalities (graphs). Three different fusion  schemes, namely linear form, sequential form and  score combination form, are exploited in the algorithm. Experimental results on the DUC benchmark  datasets demonstrate the effectiveness of the proposed multi-modality learning algorithms with all  the three fusion schemes