At present, most text summary methods based on deep learning are supervised methods that require large dataset. However, large-scale multi-document summary(MDS) datasets are difficult to obtain in practical applications. This paper proposes an unsupervised MDS extraction method based on graph structure. First, we use a pre-trained language model based on BERT for sentence representation. Secondly, according to the semantic relation between sentences, the semantic relation graph of sentences is constructed. Then the centroid and maximum marginal correlation algorithm (MMR) were used to remove the redundancy of candidate summary. Finally, experiments are conducted on a Multi-news dataset to validate the effectiveness of the proposed method. The experimental results show that the method proposed in this paper achieves better results in the unsupervised domain and is comparable with some supervised models. The method proposed in this paper has wide application prospects in the field of multi-document sentiment summarization.