Most of the previous researches on multi-document summarization concentrate on factual text. How to get a summary that cover the most important topics of evaluative text and to extract the main opinion on these topics have not been solved completely yet. This paper considers both sentiment and mixture of topics of a sentence simultaneously and proposes a multi-document opinion summarization approach based on Latent Dirichlet Allocation model and semantic orientation of word. Experiments on product reviews show that this approach performs better.